BLAKE3 is a fast new cryptographic hash algorithm created
by the team behind BLAKE2. Python's BLAKE3 support uses local
SIMD extensions where available, making it Python's fastest
built-in hash algorithm on newer Intel and AMD CPUs for
non-trivial inputs.

Touched up the blake3 object interface so it's a better match
with the the "blake3" package on PyPI:
  * derive_key_context is now a (Unicode) string.
  * Added a reset() method.
  * Enforce that key, if specified, is exactly 32 bytes.
  * Error strings used in exceptions are converging with what's
    in the "blake3" package.

Also, improved the setup.py build logic:
  * It now preserves the .o files from one build to another.
  * It does time stamp comparisons, so it rebuilds the .o file
    if it's out of date.
  * It remembers if building an extension completely failed.
In short, it now behaves completely properly like a normal "make".
It never extraneously re-tries a compilation.