Add GFNI bit-transpose path

[stdpain]

The AVX512 bit transpose (bshuf_trans_bit_byte / bshuf_shuffle_bit_eightelem) peels the 8 bit-planes out of each 64-byte block with an 8-pass loop of vpcmpb-to-mask + vpsllw. On CPUs with GFNI a single vgf2p8affineqb transposes the 8x8 bit-matrix of every 64-bit lane, and vpermb (AVX512VBMI) regroups the bytes -- ~4 ops in place of ~24.

A new USEGFNI path (gated on USEAVX512 && GFNI && AVX512VBMI; every AVX512-class CPU exposing GFNI also exposes VBMI) provides drop-in, byte-identical replacements wired into bshuf_trans_bit_elem_AVX512 and bshuf_untrans_bit_elem_AVX512. Output is bit-for-bit unchanged, so the compressed format is preserved. using_GFNI() is exposed alongside the existing using_* accessors (C and Python).

Validated on Zen4 (transform is vendor-independent):

• GFNI transpose byte-identical to TRANS_BIT_8X8 (all 64 basis vectors + random);
• full round trip for elem_size in {1,2,4,8,16}, aligned and unaligned sizes;
• compressed output byte-identical (same checksum) to the non-GFNI build.

Bit-transpose kernel: AVX2 ~5, AVX512 ~8, GFNI ~33-55 GB/s (~4-6x over the AVX512 loop; AVX2->AVX512 is only ~1.5x -- same algorithm, wider registers). End-to-end bshuf_decompress_lz4 improves ~1.3-1.7x, largest on high-entropy / floating-point data where the transpose dominates decode time.

Performance

Measured on AMD Zen4 (Ryzen 9 8945HX). The transform is vendor-independent and
applies equally to Ice Lake / Sapphire Rapids.

Bit-transpose kernel (the stage this PR changes)

Throughput of the isolated bit transpose, all three code paths doing the same work:

size	AVX2	AVX512	GFNI	GFNI vs AVX512
8 KB	6.5 GB/s	9.7 GB/s	55 GB/s	5.7×
64 KB	5.5 GB/s	8.5 GB/s	38 GB/s	4.5×
256 KB	5.3 GB/s	8.2 GB/s	38 GB/s	4.7×
1 MB	5.2 GB/s	7.8 GB/s	33 GB/s	4.2×

The AVX2 → AVX512 step is only ~1.5× (same 8-pass algorithm, wider
registers); GFNI wins ~4–6× by changing the algorithm, not the width.

End-to-end bshuf_decompress_lz4 (elem_size = 8)

The bit transpose is 43–84% of decode time (highest on incompressible data,
where LZ4 is close to memcpy), so the kernel speedup carries through:

data	ratio	AVX512	GFNI	speedup
high entropy	0.995	4.98 GB/s	8.64 GB/s	1.73×
float (noisy)	1.21	4.38 GB/s	6.97 GB/s	1.59×
low entropy	7.6	3.64 GB/s	5.35 GB/s	1.47×
sorted ids	20.9	2.75 GB/s	3.62 GB/s	1.32×