Use AVX512_VPOPCNTDQ for the population-count kernels

The AVX2 kernels build a population count out of a VPSHUFB nibble lookup
plus VPSADBW and handle 256 bits per iteration. AVX512_VPOPCNTDQ provides
VPOPCNTQ, which counts all eight 64-bit lanes of a ZMM register in one
instruction, so the new kernels handle 2048 bits per iteration with four
accumulators to keep the adds off a single dependency chain.

One 8 KB container on a Xeon Gold 6548N:

    popcntSlice      179.9 ns -> 44.7 ns  (4.02x vs AVX2)
    popcntAndSlice   225.9 ns -> 68.5 ns  (3.30x vs AVX2)
    popcntOrSlice    226.9 ns -> 68.5 ns  (3.31x vs AVX2)

These back GetCardinality, computeCardinality, AndCardinality,
OrCardinality, XorCardinality, Intersects, and the cardinality half of
every bitmap-container boolean operation.

Detection goes through golang.org/x/sys/cpu, so the OSXSAVE and XCR0
opmask/ZMM checks are handled upstream and GODEBUG=cpu.avx512vpopcntdq=off
disables the path at run time. CPUs without VPOPCNTQ keep the AVX2 path.

Claude-Session: https://claude.ai/code/session_0123ePBefqPjrCvhxdwWFakj