Auto merge of #153640 - Zalathar:subchunk, r=wesleywiser

Make bitset `would_modify_words` more vectorizer-friendly



Currently this function compares a single pair of `u64` at a time, which is potentially slower than comparing multiple words before each early-exit check, especially for the large chunks used by ChunkedBitSet.

Perf shows a notable improvement in `cranelift-codegen`, which is the one benchmark that is known to stress these code paths.

- Incorporates rust-lang/rust#153759.

Author: bors