Better narrowing for enums and other types with known equality

[hauntsaninja]

Fixes#21187, fully fixes this comment #9003 (comment) (previously improved on in 1.20 narrowing changes)

This diff adds general functionality that replaces a few different pieces of ad hoc logic:

• The conservative enum handling is now generalised
• We can get rid of the special cased logic for None when narrowing contains, it is fully subsumed by more general logic
• We no longer rely on promotion behaviour to narrow numeric types correctly
• Similarly, we now narrow bytes consistently, regardless of --strict-bytes. We also no longer have to mark memoryview and bytearray as having custom equality implementations. That is, it is a long term fix for Optional bytes type not narrowed after asserting equality #20701

Co-authored-by Codex

[JukkaL]

Thanks, a few minor comments, otherwise looks good.

[JukkaL]

The above two reveal types aren't strictly consistent, but it's probably fine since mixing bytearray or memoryview objects with bytes is very rare (though it happens occasionally).

[hauntsaninja]

Agreed. This PR doesn't actually affect that behaviour, but I added this to the test case because I wanted to document it for this reason

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

rich (https://github.com/Textualize/rich)
+ rich/console.py:1542: error: Redundant cast to "Literal['left', 'center', 'right']" [redundant-cast]
cryptography (https://github.com/pyca/cryptography)
+ src/cryptography/hazmat/primitives/serialization/ssh.py:1120: error: Argument 16 to "SSHCertificate" has incompatible type "bytes | Any | memoryview[int]"; expected "bytes" [arg-type]