[Codegen] Avoid unnecessary masking in map_scatter

[Yu-Zhewen]

When folding extract_slice operations into map_scatter, the current behavior is always adding masking logic, without checking whether the extract_slice is an identity. For static shapes, this is fine because later optimizations will remove the unnecessary masking. However, for dynamic shapes, the masking code persists and adds runtime overhead. This PR therefore avoids inserting unnecessary masking at the first place.

This PR also changes the behavior of pure-tensor, identity map_scatter canonicalization from being replaced with a copy operation to simply erased. The previous use of copy operations was causing issues with LLVMGPUSelectLoweringStrategyPass.

This PR helps mitigate the VGPR spilling issue observed in #21865.

Depends on llvm/llvm-project#173356.