Add AVX-512 16-wide encoder backend with VBMI table gathers

[madanm3]

Summary

• Add astcenc-avx512: a compile-time 16-wide AVX-512 encoder, same idea as the existing AVX2 8-wide build.
• Float-table gathers use AVX-512 VBMI permutes (load the table once, permute many times), not vgatherdps. Needs AVX-512F and AVX-512VBMI. Other ISAs, including NEON, are unchanged.

Build

cmake -DCMAKE_BUILD_TYPE=Release -DASTCENC_ISA_AVX512=ON -DASTCENC_INVARIANCE=ON ..

• Needs AVX-512F + AVX-512VBMI in hardware and enabled in the OS (XCR0). Skylake-X / Cascade Lake (F without VBMI) will not run this binary.
• The CLI checks that; the library does not. MSVC is not supported for this backend (Clang-CL only).
• Universal macOS + AVX-512 is rejected.

Testing

• ctest AVX-512 unit tests (SIMD width 16, gatherf_table 16/32/64/144)
• Functional CLI (--encoder avx512)
• Small image suite vs same-machine main AVX2: PSNR 0.000 dB; 6×6 -medium ASTC bit-identical (76559cfd…)

Performance

Speedup is main AVX2 coding time / this binary (>1 = this PR is faster). Small set, 3 repeats, encodes ≥2 ms. Invariance on.

Ryzen 7 9700F — Clang 18, -j 8

Quality	Overall	4×4	6×6	8×8	12×12
fastest	1.20×	1.15×	1.20×	1.40×	1.05×
fast	1.21×	1.14×	1.23×	1.43×	1.06×
medium	1.27×	1.17×	1.32×	1.54×	1.09×
thorough	1.31×	1.18×	1.38×	1.62×	1.13×

Pinned 6×6 medium (taskset -c 4, -j 1): 0.0836 s → 0.0621 s (1.35×).

Ryzen 7 9700F — g++ 13.3, -j 8

Quality	Overall	4×4	6×6	8×8	12×12
fastest	1.15×	1.03×	1.11×	1.34×	1.11×
fast	1.16×	1.03×	1.15×	1.37×	1.10×
medium	1.20×	1.06×	1.23×	1.46×	1.10×
thorough	1.25×	1.09×	1.29×	1.53×	1.14×

Pinned 6×6 medium: 0.0796 s → 0.0635 s (1.25×).

Xeon Platinum 8380 (Ice Lake) — g++ 14.3, encode-only 20 Small PNGs, -j 8

Quality	Overall	4×4	6×6	8×8	12×12
fastest	0.94×	0.98×	1.00×	1.02×	0.71×
fast	0.95×	0.99×	1.00×	1.03×	0.73×
medium	0.95×	1.00×	1.02×	1.05×	0.70×
thorough	0.97×	1.00×	1.03×	1.07×	0.71×

Pinned 6×6 medium: 0.1239 s → 0.1214 s (1.02×). The 16-wide/VBMI win is a Zen 5 result; Ice Lake is flat at 6×6 and loses 12×12 (64-float table cap + 512-bit license).

Apple M5 — NEON vs main NEON, Apple Clang 17

Quality	Overall	4×4	6×6	8×8	12×12
fastest	0.999×	1.000×	0.997×	0.994×	0.998×
fast	1.002×	1.001×	1.006×	1.001×	1.002×
medium	1.001×	1.000×	1.002×	1.001×	0.999×
thorough	1.000×	1.000×	0.997×	1.004×	1.001×

Pinned 6×6 medium: 0.0542 s vs 0.0541 s. Same ASTC hashes as x86 (76559cfd… / 6412a7ae…). ARM is unchanged.

Commits

1. ISA / CMake / codec hoists / CLI CPU check
2. README, changelog, Building.md, weekly CI avx512 package, test_simd, Python --encoder avx512