[llvm] [SLP]Improve/fix subvectors in gather/buildvector nodes handling (PR #104144)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

> This reproduces with i686 and x86_64, but not on aarch64.

In this one, SLP is producing the illegal instruction:
```ll
%51 = call <8 x i16> @llvm.vector.insert.v8i16.v4i32(<8 x i16> %50, <4 x i32> %41, i64 4)
```

https://github.com/llvm/llvm-project/pull/104144