* Fix buffer protocol implementation

According to the buffer protocol, `ndim` is a _required_ field [1], and
should always be set correctly. Additionally, `shape` should be set if
flags includes `PyBUF_ND` or higher [2]. The current implementation only
set those fields if flags was `PyBUF_STRIDES`.

[1] https://docs.python.org/3/c-api/buffer.html#request-independent-fields
[2] https://docs.python.org/3/c-api/buffer.html#shape-strides-suboffsets

* Apply suggestions from review

* Obey contiguity requests for buffer protocol

If a contiguous buffer is requested, and the underlying buffer isn't,
then that should raise. This matches NumPy behaviour if you do something
like:
```
struct.unpack_from('5d', np.arange(20.0)[::4])  # Raises for contiguity
```

Also, if a buffer is contiguous, then it can masquerade as a
less-complex buffer, either by dropping strides, or even pretending to
be 1D. This matches NumPy behaviour if you do something like:
```
a = np.full((3, 5), 30.0)
struct.unpack_from('15d', a)  # --> Produces 1D tuple from 2D buffer.
```

* Handle review comments

* Test buffer protocol against NumPy

* Also check PyBUF_FORMAT results