fix(markers): normalize extras/dependency_groups membership values at parse time

[r266-tech]

Problem

_normalize_extras (the parse-time normalizer run in Marker.__init__ / Requirement.__init__) canonicalizes only the singular extra == "X" form. The plural, set-valued membership forms "X" in extras and "X" in dependency_groups are left un-normalized at parse time, so a marker's __str__ / __eq__ / __hash__ disagree with its evaluate() result — and with PEP 685 (extras) / PEP 735 (dependency-groups), which mandate PEP 503 normalization of these names.

>>> from packaging.markers import Marker
>>> Marker('"Foo" in extras') == Marker('"foo" in extras')
False                              # str/hash differ ...
>>> Marker('"Foo" in extras').evaluate({"extras": {"foo"}})
True
>>> Marker('"foo" in extras').evaluate({"extras": {"foo"}})
True                               # ... yet they evaluate identically
>>> Marker('extra == "Foo"') == Marker('extra == "foo"')
True                               # the singular form is already normalized

evaluate() is consistent because _normalize canonicalizes both operands whenever key in MARKERS_ALLOWING_SET; only the parse-time path was missing the plural case. The practical consequence is that equal extras/dependency-groups markers fail to dedup in sets/dicts and compare unequal in requirement/lock-file processing.

Fix

Extend _normalize_extras to canonicalize the string-literal operand when the other operand is a Variable named extras/dependency_groups, mirroring the existing extra guards. The membership variable is always the right-hand operand (Value in Variable), so a single branch covers both in and not in. No operator check is needed because _normalize already normalizes these keys regardless of operator, so parse-time and eval-time now agree.

This is a follow-up completing the plural-membership gap left by the recent PEP 685 normalization work (#1246 nested singular extra, #1278 Requirement extras).

Tests

Added test_membership_value_str_normalization and ..._negated (parametrized over extras / dependency_groups) asserting str, ==, and hash consistency for both in and not in. Full tests/test_markers.py (2284 tests) passes; ruff check and ruff format are clean.