[ty] Patch typeshed mapping key annotations during sync (#25831)
## Summary
Updated our typeshed sync to:
- patch `dict.get` and `Mapping.get` to accept `object` keys
- patch `dict_keys.isdisjoint` and `dict_items.isdisjoint` to accept
`object`
- apply the typeshed patches at the end of the sync in a standalone
commit
The workflow pushes the completed, unpatched sync first, so a patch
conflict can be fixed from the unfinished sync branch without rerunning
the full workflow. The currently vendored stubs include the patched
signatures immediately.
Used a typeshed-patching approach similar to mypy, with a directory of
patch files that we apply in order. If a patch fails to apply, the final
sync job fails while leaving the unpatched sync branch available for
manual repair.
Closes https://github.com/astral-sh/ty/issues/2374
## Rationale
These methods can safely be called with any key without erroring, so the
current annotations are too strict and cause false positives. See
https://github.com/python/typeshed/issues/15271 for upstream discussion
in typeshed.
It's reasonable to want a diagnostic in cases when the `.get()` call is
known to return `None`, or the `isdisjoint` call is known to be always
true; switching to `object` loses that. In the short term we could
provide this via dedicated rule. Longer-term there is discussion
upstream about an `Overlapping[...]` annotation form that would allow
passing any type that is not disjoint with the annotated type.
There are other methods mentioned in that typeshed issue that in
principle we might want to patch for similar reasons, but I'm preferring
to keep the patch-set smaller and start with just the methods we've seen
come up.
## Validation
- Added mdtest regression coverage for every `get` overload form
accepting `object`.
- Ran the focused `call/builtins.md` mdtest.
- Ran the full `ty_python_semantic` test suite: 558 passed.
- Applied the exact workflow `git apply` command to the current
post-docstring, post-format vendored baseline and verified the results
byte-for-byte.
- Verified that applying the patches a second time fails rather than
silently succeeding.
- Ran file-scoped `prek` hooks, including GitHub workflow validation.
Ecosystem results are mostly the expected false-positive removals, plus
some downstream additions where allowing the `.get()` call gives a
precise type instead of `Unknown`, causing downstream diagnostics.
This change does mean that subclass overrides of `Mapping.get` may now
emit Liskov violations if they don't accept `object` -- I think this is
an unavoidable part of this change?
