Investigate performance of integer comparisons for mask reductions

[gnzlbg]

Currently, all / any mask vector reductions are implemented as:

• x86/x86_64:
  • mmx: using _mm_movemask_pi8
  • sse2: using _mm_movemask_epi8
  • avx: using _mm256_testc_si256, _mm256_testz_si256
• arm/aarch64:
  • neon: using vpmin, vpmax
• other architectures call the llvm.experimental.vector.reduce.{and, or} intrinsics.

It might be potentially much better to just cast all vectors <= 128-bit wide to their corresponding integer and do a != 0 for any, and a == iX::max_value() for the all reduction.

For 256-bit and 512-bit wide vectors we could cast them into an [u128; N] array and perform the integer comparisons against that.