[LoopVectorize] Refine runtime memory check costs when there is an outer loop

[david-arm]

We would like to backport #76034 to fix a major performance regression when using PGO to build certain applications/benchmarks. Specifically, it fixes a 25% regression in the SPEC2017 x264 benchmark.

[nikic]

/cherry-pick 962fbaf

[llvmbot]

/pull-request #79813

[tstellar]

Merged: bdaf16d