viable/strict/1780442474: Separate value index expressions from indexing (#185853)

Introduce value_expr as the dtype-honoring form of index_expr for symbolic

expressions that participate in tensor value computation. A pass classifies

index_expr nodes as indexing or value uses by walking backward from value sinks,

then converts value uses to value_expr in-place. On Triton, value_expr

patches the kernel index dtype and casts the result so arithmetic runs at

the correct width (e.g. int64 for arange(dtype=int64)).

In the future, this should also allow us to no-longer have a "per-kernel" index dtype, but correctly strength reduce particular indexing expressions based on which tensors they are used in.

See, more analysis here https://docs.google.com/document/d/1AdAUkhzd_3qL2SH4kV9AWj2e55yVrNFqbiQP91beQA0/edit?tab=t.0.

Fixes multiple bugs. Note - most of LOC here is just test changes. happy to simplify tests etc as needed.

Authored with Claude.

Pull Request resolved: #185853

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