[llvm] GlobalISel: Don't use GISelMatchGenericTypes for floating point opcodes (PR #212814)

[Petar Avramovic via llvm-commits]

petar-avramovic wrote:

Applied to all fp atomics, G_ATOMICRMW_XCHG is fine with scalar check. amdgpu only had issue for f16 vs bf16 atomic fadd

https://github.com/llvm/llvm-project/pull/212814