[X86] Loop-invariant operand defeats vector lowering of i256/i512 CTPOP and logic ops

Noticed with `_BitInt(512)` in a Hamming-distance loop. #163788 fixed the case where both operands are loaded; the scalar chain comes back when one of them is loop-invariant.

https://godbolt.org/z/G9jdnr8eb

```c
typedef unsigned _BitInt(512) u512;

void inv(const u512 *db, u512 q, int n, int *o) { for (int i=0;i

Start at X86::mayFoldIntoVector and inspect how CTPOP and logic operands are admitted, then reproduce the inv, constant-mask, sparse-mask, and volatile cases with the provided IR or llc runs. Add coverage under llvm/test/CodeGen/X86, including the generated invariant, GPR-resident, and volatile cases. Done means the invariant loop vectorizes without regressing the three scalar cases.

Written by the indexing model from the issue text.