Fix metadata dependency cache leaking across the MRO

[Sanjays2402]

Closes #1459

Summary

MetadataDependent.get_inherited_dependencies() memoized into a plain class
attribute and read it back with ordinary attribute lookup, which walks the MRO.
Resolving a bare CSTTransformer/CSTVisitor during a metadata pass therefore
cached an empty frozenset onto the base class, and every subclass that had not
yet computed its own cache inherited that empty set — raising a spurious
METADATA_DEPENDENCIES KeyError even though the subclass declared its
provider correctly. The poisoning is process-wide and surfaces far from its
cause.

The fix reads only the class's own __dict__ entry, so a base-class cache can
never be inherited. Memoization and its per-class semantics are unchanged.

Test Plan

Added test_bare_base_class_does_not_poison_inherited_metadata in
libcst/metadata/tests/test_metadata_provider.py, beside the existing
test_inherited_metadata: it resolves a bare CSTTransformer first, then a
subclass declaring a provider. Stash-verified — the test fails without the
source change (the exact reported KeyError) and passes with it.
libcst/metadata/tests/ is 174 passed; the 3 errors are the pre-existing pyre
test_type_inference_provider cases that need the pyre binary.

This change was prepared with AI assistance; the regression test was run locally
and fails without the fix.

[Sanjays2402]

Ping on this one — CI is blocked on the first-run approval gate. Note this supersedes #1466 (now closed), which fixed the same issue with a larger diff. Happy to adjust if you'd prefer a different approach to #1459.