* stl.h: propagate return value policies to type-specific casters

Return value policies for containers like those handled in in 'stl.h'
are currently broken.

The problem is that detail::return_value_policy_override<C>::policy()
always returns 'move' when given a non-pointer/reference type, e.g.
'std::vector<...>'.

This is sensible behavior for custom types that are exposed via
'py::class_<>', but it does not make sense for types that are handled by
other type casters (STL containers, Eigen matrices, etc.).

This commit changes the behavior so that
detail::return_value_policy_override only becomes active when the type
caster derives from type_caster_generic.

Furthermore, the override logic is called recursively in STL type
casters to enable key/value-specific behavior.

Switching deprecated Thread Local Storage (TLS) usage in Python 3.7 to Thread Specific Storage (TSS) (#1454)

* Switching deprecated Thread Local Storage (TLS) usage in Python 3.7 to Thread Specific Storage (TSS)

* Changing Python version from 3.6 to 3.7 for Travis CI, to match brew's version of Python 3

* Introducing PYBIND11_ macros to switch between TLS and TSS API

* Silence GCC8's -Wcast-function-type.

See https://bugs.python.org/issue33012 and PRs linked therein.

It is useful not only to remember the python libs and includes but
also the interpreter version in cache.

If users call pybind11 throught `add_subdirectories` they will
otherwise have no access to the selected interpreter version.
The interpreter version is useful for downstream projects, e.g.
to select default `lib/pythonX.Y/site-packages/` install paths.

Add missing python3 distutils on Debian "buster".

If an exception is thrown during module initialization, the
error_already_set destructor will try to call `get_internals()` *after*
setting Python's error indicator, resulting in a `SystemError: ...
returned with an error set`.

Fix that by temporarily stashing away the error indicator in the
destructor.

Found via `codespell`

Fixes #1401

Fix a segfault when creating a 0-dimension, c-strides array.

As reported in #1349, clang before 3.5 can segfault on a function-local
variable referenced inside a lambda.  This moves the function-local
static into a separate function that the lambda can invoke to avoid the
issue.

Fixes #1349

This fixes the test code on big-endian architectures: the array support
(PR #832) had hard-coded the little-endian '<' but we need to use '>' on
big-endian architectures.

Current MSVC generates totally bizarre errors:

    error C2884: 'pybind11::detail::_': introduced by using-declaration
    conflicts with local function 'pybind11::detail::_'

which makes no sense (since the supposed "conflict" is the function
itself).  Work around it by `using namespace detail;` instead (which
also lets us drop a bunch of other `detail::` qualifications, so isn't
actually a bad thing).

This updates the `py::init` constructors to only use brace
initialization for aggregate initiailization if there is no constructor
with the given arguments.

This, in particular, fixes the regression in #1247 where the presence of
a `std::initializer_list<T>` constructor started being invoked for
constructor invocations in 2.2 even when there was a specific
constructor of the desired type.

The added test case demonstrates: without this change, it fails to
compile because the `.def(py::init<std::vector<int>>())` constructor
tries to invoke the `T(std::initializer_list<std::vector<int>>)`
constructor rather than the `T(std::vector<int>)` constructor.

By only using `new T{...}`-style construction when a `T(...)`
constructor doesn't exist, we should bypass this by while still allowing
`py::init<...>` to be used for aggregate type initialization (since such
types, by definition, don't have a user-declared constructor).

This fixes #1251 (patient vector grows without bounds) for the 2.2.2
branch by checking that the vector doesn't already have the given
patient.

This is a little less elegant than the same fix for `master` (which
changes the patients `vector` to an `unordered_set`), but that requires
an internals layout change, which this approach avoids.

* Fix segfault when reloading interpreter with external modules

When embedding the interpreter and loading external modules in that
embedded interpreter, the external module correctly shares its
internals_ptr with the one in the embedded interpreter.  When the
interpreter is shut down, however, only the `internals_ptr` local to
the embedded code is actually reset to nullptr: the external module
remains set.

The result is that loading an external pybind11 module, letting the
interpreter go through a finalize/initialize, then attempting to use
something in the external module fails because this external module is
still trying to use the old (destroyed) internals.  This causes
undefined behaviour (typically a segfault).

This commit fixes it by adding a level of indirection in the internals
path, converting the local internals variable to `internals **` instead
of `internals *`.  With this change, we can detect a stale internals
pointer and reload the internals pointer (either from a capsule or by
creating a new internals instance).

(No issue number: this was reported on gitter by @henryiii and @aoloe).

Fix return from `std::map` bindings to `__delitem__`: we should be returning `void`, not an iterator.

Also adds a test for map item deletion.

Found via `codespell`

The anonymous struct nested in a union triggers a -Wnested-anon-type
warning ("anonymous types declared in an anonymous union are an
extension") under clang (#1204).  This names the struct and defines it
out of the definition of `instance` to get around to warning (and makes
the code slightly simpler).

- UPDATEIFCOPY is deprecated, replaced with similar (but not identical)
  WRITEBACKIFCOPY; trying to access the flag causes a deprecation
  warning under numpy 1.14, so just check the new flag there.
- Numpy `repr` formatting of floats changed in 1.14.0 to `[1., 2., 3.]`
  instead of the pre-1.14 `[ 1.,  2.,  3.]`.  Updated the tests to
  check for equality with the `repr(...)` value rather than the
  hard-coded (and now version-dependent) string representation.

When using the mixed position + vararg path, pybind over inc_ref's
the vararg positions. Printing the ref_count() of `item` before
and after this change you see:

Before change:

```
refcount of item before assign 3
refcount of item after assign 5
```

After change
```
refcount of item before assign 3
refcount of item after assign 4
```

The `py::args` or `py::kwargs` arguments aren't properly referenced
when added to the function_call arguments list: their reference counts
drop to zero if the first (non-converting) function call fails, which
means they might be cleaned up before the second pass call runs.

This commit adds a couple of extra `object`s to the `function_call`
where we can stash a reference to them when needed to tie their
lifetime to the function_call object's lifetime.

(Credit to YannickJadoul for catching and proposing a fix in #1223).