x86 backend issue in AVX-512 code when generating constant -1 in zmm reg

Observed in Clang 22.1.0, but I can reproduce this as far back as Clang 4.0, so it's been there forever.

Small repro on Compiler Explorer: https://godbolt.org/z/anceexcT7

The `set1_epi32(-1)` turns into `vpternlogd zmm1, zmm1, zmm1, 255`, presumably adapted from the older "ones' idiom" using `pcmpeq* reg, reg`. The issue being that the `pcmpeq` version is a dependency-breaking idiom on Intel/AMD cores, but the `vpternlogd` version is not, at least not on the Zen 4 machine I tested this on.

The original case I encountered had this `vpternlogd` inside a hot loop where the iterations are supposed to be independent, and the `vpternlogd` introduced a loop-carried dependency that decreased throughput by about 34%.

This seems iffy to do in general and should probably just use a constant load or similar on AVX-512 targets.

Start by running the linked Compiler Explorer repro and comparing the AVX-512 code generated for set1_epi32(-1), especially the vpternlogd instruction. Trace the x86 backend's constant-generation decision and verify that the completed change removes the unwanted dependency without regressing the generated code.

Written by the indexing model from the issue text.