[llvm] [AMDGPU] Do not treat bitcast across FP types as canonicality-preserving (PR #203560)

[Arseniy Obolenskiy via llvm-commits]

aobolensk wrote:

> The way this function handles casts, truncates etc has never made much sense to me. I feel like the API probably needs to change to pass in the original FP type so that recursive calls can make queries like "is this i32 value canonical _when considered as a bitcast of v2f16_?".

Passed the FP type to `isCanonicalized` check

https://github.com/llvm/llvm-project/pull/203560