[llvm] [AMDGPU][SIPreEmitPeephole] Fold consecutive `s_cselect_b32`s (PR #183357)

[Zach Goldthorpe via llvm-commits]

zGoldthorpe wrote:

Sorry to revive a dead PR. The main reason I opted to perform this optimisation so late in the compilation pipeline is because prematurely fusing 32-bit selects limits other optimisations; for example.

It seems that fusing earlier was attempted before and reverted in [`c0581f7d`](https://github.com/llvm/llvm-project/commit/c0581f7d)) due to it causing regressions.

At the very least, fusing this late won't risk obstructing e.g. scheduling or regalloc.
Is there a reasonable middle-ground place for this optimisation?

https://github.com/llvm/llvm-project/pull/183357