[llvm] GlobalISel/Combiner: Fix building G_CONSTANT for fp extended LLT (PR #196397)

[Petar Avramovic via llvm-commits]

petar-avramovic wrote:

With extended LLTs many of the opcodes that had Int/FP type encoded in opcode name probably loose usefulness. 
For AMDGPU G_CONSTANT and G_FCONSTANT are really the same thing (move constant bag of bits to register).
Now I am trying to make minimal changes in common code to enable bfloat since extended LLTs finally got merged.
Now G_CONSTANT and G_FCONSTANT have explicit type checks with extended LLTs,, but the problem is this combine happens very late in regbankcombiner so I think it needs to keep same opcode.
For AArch64 maybe lower G_FCONSTANT to G_CONSTANT, might need bitcast, in legalizer if it affects performance?

https://github.com/llvm/llvm-project/pull/196397