[Codegen] Simplify bitcast folding with bufferized load

[Yu-Zhewen]

Currently, #22672 needs to match a long chain of operations to fold bitcasts into bufferized tensor loads:

1. iree_codegen.load_from_buffer
2. amdgpu.fat_raw_buffer_cast
3. hal.interface.binding.subspan

However, this pattern matching is quite fragile.

A more robust solution is to push the bitcast directly onto the source memref of the iree_codegen.load_from_buffer op.

However, the current memref.reinterpret_cast operation doesn't support casting element types. Use memref.view, but it requires the memref to be contiguous.

Therefore, a cleaner solution would be to introduce a dedicated memref.bitcast (or similar).