Towards #5479

Unlike other comparing algorithms, mismatch vectorization can be expanded towards arbitrary element size easily, without even adding separately-compiled code. This is currently possible for clang-cl only, and it requires defaulted operator==, and trivially-comparable types, and unique object representation.

The mismatch result can be divided by element size in the header, thus truncating any offset into the middle, giving the element index. The division is likely to be optimized to large multiplication by the compiler.

Rather than using just __std_mismatch_1, attempt to use as large element mismatch as possible, for the following reasons:

• Primarily, optimal tail processing. The reason to have _2 / _4 / _8 versions in the first place.
• Also potentially smaller divisor is more likely to be optimized to the multiplication by the compiler. Though I haven't checked whether the compiler can get rid of power-of-two factor on its own; probably it can.

lexicographical_compare vectorization cannot be expanded this way, as it requires a different flavor of trivial comparability.

Benchmark results