fix(CuTeDSL): preserve wrapped signatures for type checkers

[iwzbi]

Summary

@dsl_user_op decorates every user-facing CuTe DSL wrapper (local_tile, make_tiled_mma, copy, gemm, etc.), but its annotation was Callable[..., Any] -> Callable[..., Any], which erases the wrapped callable's signature for static analyzers.

Pyright/basedpyright (and language servers built on them) therefore see only (*args, **kwargs). In practice this means:

• Editors show no parameter names in signature help: cute.local_tile( yields no tiler/coord hints;
• Bogus keyword calls compile silently (e.g. cute.local_tile(mA, xxxxx=...) passes type checking);
• Downstream projects that write their own @dsl_user_op wrappers (vLLM, SGLang, FlashInfer) are affected the same way.

Change

Make the decorator signature-preserving with a TypeVar bound to Callable, the pattern recommended by pyright's documentation:

_FuncT = TypeVar("_FuncT", bound=Callable[..., Any])

def dsl_user_op(opFunc: _FuncT) -> _FuncT:
    ...
    return wrapper  # type: ignore[return-value]

Runtime behavior is unchanged (functools.wraps was already applied to the wrapper). # type: ignore[return-value] documents that the inner wrapper intentionally has a generic signature but must satisfy _FuncT.

Verification

• python -m py_compile op.py — syntax OK.
• pyright, after patch: cute.local_tile(mA, tiler=(128, 64), coord=(0, 0)) type-checks; cute.local_tile(mA, xxxxx=(128, 64)) is rejected with No parameter named "xxxxx", and Arguments missing for parameters "tiler", "coord" is reported for incomplete calls.

Fixes editor/signature-help blindness for the entire public CuTe DSL surface.