trunk/7567e9484a51d8a1cf14ad6d919631aed013aab0: [inductor] Fix index_expr CSE variable dtype (#182872)

ks* kernel args (symbolic scalars like alpha in torch.add) are always

int64 in the Triton signature (per _decide_tl_dtype), but index_expr

was setting the CSE variable dtype to the kernel's index_dtype which

can be int32 for small tensors. This caused two issues:

1. When a ks* arg needed casting (e.g., int64 -> float for math.sqrt),
   
   the CSE variable's wrong dtype could cause the cast to be skipped.


2. runtime_triton_dtype_assert would fire incorrect static_assert
   
   checks comparing against int32 when the actual type was int64.

Fix: use torch.promote_types(dtype, index_dtype) so that int32

requests follow the kernel's index dtype, but int64 requests are never

downcast to int32 (which would overflow values like randint's high

bound). Also skip redundant dtype/shape assertions in

CSEProxy._default for ops like index_expr that manage their own

checking internally.

This enables test_op_dtype_propagation to run with dynamic=True.

Authored with Claude.

Pull Request resolved: #182872

Approved by: https://github.com/ezyang, https://github.com/liqiangxl