Fix out-of-bounds prefetch pointer arithmetic

[russellromney]

What changed

Use wrapping_add when computing the three x86 SIMD prefetch addresses.

Why

The 16 KiB prefetch distance routinely places the address beyond the input allocation. The prefetch instruction itself is non-faulting, but Rust's ptr::add requires the computed pointer to remain within or one past the allocation even when it is not dereferenced. wrapping_add preserves the intended machine behavior without creating an invalid in-bounds-derived pointer.

Validation

• cargo fmt --all -- --check
• cargo test --target x86_64-apple-darwin prefetch_pointer_may_wrap_beyond_a_short_slice
• existing argmin differential test

The regression helper/test gives provenance-aware tools such as Miri a direct path that computes a deliberately far-out prefetch pointer from a one-byte slice without invoking a SIMD prefetch instruction.