Make warn-unreachable understand exception-swallowing contextmanagers

[Michael0x2a]

This pull request fixes #7214: it makes mypy treat any contextmanagers where the __exit__ returns bool or Literal[True] as ones that can potentially swallow exceptions.

Contextmanagers that return Any, Optional[bool], None, or Literal[False] continue to be treated as non-exception-swallowing ones.

This distinction helps the --warn-unreachable flag do the right thing in this example program:

from contextlib import suppress

def should_warn() -> str:
    with contextlib.suppress(IndexError):
        return ["a", "b", "c"][0]

def should_not_warn() -> str:
    with open("foo.txt") as f:
        return "blah"

This pull request needs the typeshed changes I made in python/typeshed#3179 before the tests and the mypy self-check will pass. Once that one gets merged, I'll update typeshed and rebase this pull request.

[ilevkivskyi]

I am going to close/re-open to trigger builds.

[Michael0x2a]

FYI, the tests for this PR aren't going to pass -- python/typeshed#3179 needs to be landed first.

[ilevkivskyi]

Oh, I somehow thought it was already merged. We can make then a dedicated sync when this will be ready. Also there is now a bonus self-check failure (easy to fix however).

[JelleZijlstra]

I merged the typeshed PR.

[ilevkivskyi]

Thanks! Looks good, I have several small comments.

[ilevkivskyi]

Are these rules/conventions documented somewhere? I think they should be, otherwise we will just forget them (either in mypy docs, btw do we have a section about context managers?, or in typeshed). Maybe also add a link here.

[Michael0x2a]

I did add some notes about these conventions to the typeshed contribution page (at the bottom of the conventions section), but this conventions currently aren't otherwise documented.

We do mention context managers once on the Protocols page -- I'll add a note about this there in a follow-up PR.

[Michael0x2a]

As an FYI, I ended up making one more change to this PR: basically, I made it so we no longer assume context managers suppress exceptions when __exit__ returns bool when strict-optional mode is disabled. I found while playing around with open(...) that we didn't necessarily preserve whether __exit__ returns bool vs Optional[bool] in that mode, so decided it'd probably be safer to just give up and bias towards retaining the old behavior.

LMK if you see a better way of handling this or if handled the no-strict-optional check incorrectly, and I can submit a follow-up PR to fix.

[ilevkivskyi]

We do mention context managers once on the Protocols page -- I'll add a note about this there in a follow-up PR.

Thanks! Sounds good to me.