[llvm] [SLP]Drop auxiliary subtrees to pass the VF=2 instruction-count veto (PR #224108)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> LGTM. Though if I understand correctly, if we have 4 different splat values, but one of the user nodes gets trimmed, we don't recreate `SplatGatheredScalarsRoots` with VF=3 and dropping that one value?

Yes, but same problem exists for cost-based trimming. I don't think it will expose significant perf regressions, but let's see

https://github.com/llvm/llvm-project/pull/224108