Treat Protocol bases as ABCMeta-providing when detecting abstract cla…

[mikeleppane]

Fixes #2797

What

A @final class that inherits collections.abc.Sequence (or Mapping) without
implementing its abstract methods is now flagged, matching mypy, pyright, and ty.

from typing import final
from collections.abc import Sequence

@final
class A(Sequence[int]):  # before: no error
    ...                  # after:  Final class `A` cannot have unimplemented
                         #         abstract members: `__getitem__`, `__len__`

Instantiating such a class (A()) and the opt-in implicit-abstract-class
diagnostic now fire correctly too.

Why

In typeshed the small collections.abc ABCs (Reversible, Collection, …) are
declared Protocol, and Protocol is a special form (Protocol: _SpecialForm)
that carries no metaclass. pyrefly's extends_abc heuristic only recognized an
abc.ABC base or an ABCMeta metaclass, so it never saw the _ProtocolMeta(ABCMeta)
that a protocol subclass actually gets at runtime. The composite Sequence — a
plain class whose only bases are those protocols — was therefore misclassified as
non-abstract and skipped, so its @final subclasses inherited no abstract
obligation. (Direct protocol bases like class B(Collection, Reversible) already
worked; the gap was one level of indirection away.) mypy, pyright, and ty are the
conformance bar and all three error here.

How

• pyrefly/lib/alt/class/class_metadata.rs — extends_abc now treats an
  is_protocol() base as ABCMeta-providing (a protocol base ⇒ _ProtocolMeta ⇒
  ABCMeta-derived metaclass). This is the whole fix for the reported bug.
• Same file — calculate_abstract_members short-circuits for is_typed_dict()
  classes. Recognizing Mapping as abstract otherwise leaked through the
  synthesized TypedDictFallback(Mapping, metaclass=ABCMeta) base and made every
  TypedDict look abstract; a TypedDict is a concrete dict and never is. Caught by
  mypy_primer on pandas-stubs, not the unit suite.
• Deliberately untouched: the two abstract-member gates (already correct given a
  correct extends_abc) and the truthiness-redundancy consumer of extends_abc in
  expr.rs (widening it only suppresses more "always true" warnings on abstract
  types — conservative).

Test plan

Test (pyrefly/lib/test/abstract_methods.rs)	Protects
test_final_class_over_sequence	the repro: @final over Sequence errors
test_abstract_through_protocol_bases	instantiation, Mapping, and a user Protocol reached via a plain intermediate
test_implicit_abstract_through_sequence	opt-in implicit-abstract diagnostic on Sequence subclasses
test_concrete_subclass_of_sequence_is_clean	implementing the members stays clean — no over-reporting
test_final_class_direct_protocol_bases_still_errors	the pre-existing direct-protocol path is unchanged
test_typed_dict_is_not_abstract	TypedDicts stay concrete (regression guard)

• Full lib suite: 5340 passed, 0 failed.
• Conformance: pass, no generated changes.
• mypy_primer, wide scope (~37 projects incl. pandas-stubs, pandas, sympy, xarray): zero diff.
• format + lint: clean.