Narrow the remaining tuple type for sequence patterns with several refutable items

[ekanshul]

Fixes #21968

For a fixed-length tuple subject, visit_sequence_pattern only narrowed the remaining type when at most one subpattern could fail to match (#20896). With two or more refutable subpatterns, e.g. case True, True: on a tuple[bool, bool], the rest type stayed the whole tuple, so match statements over tuples of bools, enums or literals never became exhaustive: --enable-error-code exhaustive-match reported an unhandled case for the full tuple[bool, bool], and functions relying on the match to return got "Missing return statement".

The values that don't match a sequence pattern are the ones where some item fails to match while every item before it matches, so the rest type is now the union of one tuple per refutable position:

tuple[A, B] - tuple[a, b] == tuple[A - a, B] | tuple[a, B - b]

Positions whose subpattern always matches (an empty rest) keep their type, and the previous single-position behaviour falls out as the special case with one refutable item. The union goes through make_simplified_union, and a subject that already is a union of tuples is handled per item by the existing union branch, so the example from the issue now narrows to tuple[Literal[False], Literal[False]] before its last case and to Never after it.

One existing expectation changes: in testMatchSequencePatternNegativeNarrowing, the rest of case (1, "a") on tuple[Literal[1, 2], Literal["a", "b"]] is now tuple[Literal[2], Literal['a'] | Literal['b']] | tuple[Literal[1], Literal['b']] instead of the unnarrowed tuple.

New tests cover the revealed rest types for two and three positions (including a wildcard in the middle and a subpattern that can't match at all), narrowing of the individual names for a match a, b: subject, and exhaustiveness for tuples of bools and enums, including the more precise unhandled-case message.

🤖 Generated with Claude Code

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅