More efficient (fixed-format) serialization

[ilevkivskyi]

This makes deserialization ~2.5x faster compared to orjson. This is fully functional, but still PoC in terms of distribution logic. Some comments:

• In you want to try this in compiled mode, simply install mypy using MYPY_USE_MYPYC=1, this will install the extension automatically (for now)
• If you want to just play with the extension, or use it in interpreted mode, use pip install mypyc/lib-rt
• I translated (de-)serialization logic from JSON methods almost verbatim (including comments)
• This may be still not the most efficient way to do this, but I wanted to write something simple, that probably still gets us 90% there in terms of performance. I am still open to suggestions however
• Please forgive me if the PR looks not very polished, I feel tired, but needed some kind of closure on this :-)

Some technical notes:

• The huge try/except import blob in mypy/cache.py is temporary, it is needed for now to be able to run tests without installing mypy itself (only with test-requirements.txt).
• There is certain asymmetry with read/write for literals, this is intentional because we allow complex and/or None in some cases, but not in other cases.
• General convention is that during deserialization the type/symbol marker is consumer by the caller (except for MypyFile, which is special). There is no convention for few classes that are not types/symbols.
• I add new primitive type for native_internal.Buffer (and possible more type in future from native) for better/automatic method call specializations. If this feels wrong/risky, I can convert this to a more ad-hoc logic in transform_call_expr()

[ilevkivskyi]

Oh, hm I completely forgot 32-bit architectures are still a thing. Probably an easy fix, not sure I will have time to take a look at it today, but it should not stop any reviewers from reviewing.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅