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

[Matt Arsenault via llvm-commits]

================

----------------
arsenm wrote:

Can you avoid using a kernel for this, it only multiplies the number of instructions that aren't relevant to the test 

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