Optimize AVX(1) VPSHUFB by loading into a register

instead of repeatedly reading it from memory

AsmVectorEquals16-32   2.208µ ± 2%   1.890µ ± 2%  -14.42% (p=0.000 n=10)

Bugfix: we accidentally did signed comparisons instead of unsigned

Add VectorNotEquals APIs

Add bounds checking before calling into ASM

This avoids memory corruption for bad calls

Avoid BM2 in the AVX(1) implementation

I could've also generated all options for with/without BMI2/AVX2, but I
simply went for having the fastest option require AVX2+BMI2 and use
AVX(1) as the fallback.

Check cpuflag for BMI2

Found out the hard way I shouldn't assume that having AVX doesn't imply
having BMI2.

Add IsNaN methods for floats

Adding a testcase for NaNs revealed that didn't work correctly

Enable AVX implementations for 64 bit width inputs

The XMM registers are 128 bits, so that's only 2 rows per round. I
implemented merging 4x 2 bits together so we can write back a byte. For
that I increased the number of rounds for 64 bit inputs to 4 (instead of
the default 2).

Call VZEROALL for a 30% benchmark improvement

Call VZEROALL before leaving any ASM function.

I didn't expect this large a speedup, so I'm a little suspicious.