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

[Martin Storsjö via llvm-commits]

mstorsjo wrote:

Second reproducer:
```c
int a(char *b, int c) {
  int d, e, f = d = 0;
  for (; d < 3; d++) {
    e = 0;
    for (; e < 8; e++)
      f += -b[e] - b[e + c] >> 31;
    b += c;
  }
  return f;
}
```
This reproduces with i686 and x86_64, but not on aarch64.

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