[llvm] [SLP] Enable full non-power-of-2 vectorization by default (PR #199450)

[David Sherwood via llvm-commits]

david-arm wrote:

FYI, I tested out this PR with SPEC2017 on a graviton 4 machine and I'm seeing quite a few performance regressions with an overall drop of 0.5% geomean. For example, I see significant regressions in imagick and leela.

https://github.com/llvm/llvm-project/pull/199450