Two orthogonal but related changes:

• Exclude AVX2 search path for 8-bit elements (but not for find_end). Address concern in Vectorize search for 32-bit and 64-bit elements, also improve 8-bit and 16-bit vectorization #5484 (comment)
• Add SSE4.2 paths for 32-bit and 64-bit elements. They are expectedly about 2 times slower than AVX2 counterparts, which is still an improvement over the scalar (except for evil cases)

Nothing more to tell about the 8-bit AVX2 search path removal. Some details on SSE4,2 path for 32-bit and 64-bit elements:

• Traits are separate for AVX2 and SSE4.2, like algorithms in _Sorting or bitset algorithms
• Unlike _Sorting, BSF and BSR are traits members. In _Sorting, these functions were called infrequently, so they were not worth differentiating for AVX2/SSE4.2. Here we use them in tightest loops, so the performance of better ones may make a difference. Beware that CodeQL comments are harder to verify here
• The tricky part is making sure no beyond SSE4.2 instructions are used in SSE4.2 path, but different types help a lot, and tzcnt/lzcnt were verified manually.
• _Size_bytes_1 is not passed to extracted functions deliberately. In case of non-inline functions it looks more optimal to compute again than to pass 5th parameter and start using stack on x64. in case of inline functions the compiler will combine the same way computed variables if needed.
• Originally, I've planned to remove find_end cmpestrm path, and always use cmpeq in find_end. The Vectorize search for 32-bit and 64-bit elements, also improve 8-bit and 16-bit vectorization #5484 (comment) results convinced me not to do this, and keep all existing SSE4.2 paths for 8-bit and 16-bit elements