popcnt: expand comments in popcnt_avx2_amd64.s

Document the AVX2 popcount assembly more thoroughly. Comments only -- no
instruction was changed; the build, go vet's asm checker, and the
differential tests all still pass.

- Header split into an algorithm overview (the Mula/Lemire VPSHUFB
  nibble-lookup popcount) and a Go-assembler conventions section (operand
  order, Yn/Xn registers, FP argument layout, NOSPLIT/$0 leaf frames, and
  why VMOVDQU and VZEROUPPER are used).
- lutmask: decode the constant blob -- how the bytes map to nibble
  popcounts, why the 16-entry table is duplicated per 128-bit lane, and
  what the 0x0F mask is for.
- COUNTBLOCK: line-by-line walkthrough of the eight instructions plus the
  no-overflow reasoning; SETUP and HSUM expanded similarly.
- _popcntSliceAVX2 fully annotated as the canonical template; the
  And/Or/Xor/Mask variants note they mirror it and differ only in the
  combine op (including the VPANDN operand-order detail) and explain the
  +48 return slot.
- _hasAVX2: explain why each CPUID/XGETBV check matters.