The updated code is designed to:
- Be branch-predictor-friendly
- Be cache-friendly
- Minimize the lengths of critical paths
- Minimize slow operations (particularly multiplications)
- Minimize binary/codegen bloat

The most notable performance trick here is perhaps the precomputation & caching of the number of digits, so that we can reuse/exploit it when writing the output.

This precomputation of the exact length enables 2 further performance benefits:
- It makes `StrCat` and `StrAppend` zero-copy when only integers are passed, by avoiding intermediate `AlphaNum` entirely in those cases. If needed in the future, we can probably also make many other mixtures of non-integer types zero-copy as well.
- It avoids over-reservation of the string buffer, allowing for more strings to fit inside SSO, which will likely have further performance benefits.

There is also a side benefit of preventing `FastIntToBuffer` from writing beyond the end of the buffer, which has caused buffer overflows in the past.

The new code continues to use & extend some of the existing core tricks (such as the division-by-100 trick), as those are already efficient.

PiperOrigin-RevId: 595785531
Change-Id: Id6920e7e038fec10b2c45f213de75dc7e2cbddd1