perf: vectorize small u8 table take with AVX2

[joseph-isaacs]

Rationale

Add the x86 table-lookup implementation after the NEON implementation in #9571.

This is the upper PR in the stack: #9566 (benchmark) → #9571 (NEON) → this PR (AVX2).

Changes

• Use AVX2 VPSHUFB on x86/x86-64.
• Apply the path to u8 codes, at most 16 one-byte values, and at least 64 rows.
• Retain runtime AVX2 detection, bounds checks, and the existing general AVX2/scalar fallbacks.

CodSpeed wall-time results

Medians over 1,000 samples on the same Sapphire Rapids metal family, compared with #9566:

Build	Rows	Baseline	AVX2 table	Speedup	Time reduction
AVX2	1M	428.1 µs	61.12 µs	7.00×	85.7%
AVX2	16M	6.809 ms	1.361 ms	5.00×	80.0%
AVX-512	1M	426.5 µs	59.28 µs	7.19×	86.1%
AVX-512	16M	6.778 ms	1.358 ms	4.99×	80.0%

The AVX-512 leg is an AVX-512-enabled whole-crate build executing this AVX2 VPSHUFB kernel; there is no separate AVX-512 table kernel.

Runs: baseline, optimized.

Validation

• 3,352 vortex-array tests passed; 1 skipped
• targeted Clippy passed
• x86-64 cross-check passed
• formatting and diff checks passed

[codspeed-hq]

Merging this PR will improve performance by 16.12%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 9 improved benchmarks
✅ 1972 untouched benchmarks
🆕 6 new benchmarks
⏩ 54 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	decode_primitives[u8, (2000, 4)]	51.4 µs	41.1 µs	+25.06%
⚡	Simulation	decode_primitives[u8, (2000, 2)]	51.5 µs	41.3 µs	+24.82%
⚡	Simulation	decode_primitives[u8, (2000, 8)]	51.9 µs	41.7 µs	+24.39%
⚡	Simulation	take[small_m/shuffled/primitive/nonnull/chunks=2048/indices=64]	801.9 µs	711.9 µs	+12.64%
⚡	Simulation	decode_primitives[u8, (1000, 8)]	43.5 µs	38.8 µs	+12.16%
⚡	Simulation	decode_primitives[u8, (1000, 4)]	43.7 µs	39 µs	+12.12%
⚡	Simulation	take[small_m/shuffled/primitive/nonnull/chunks=256/indices=64]	611.3 µs	545.4 µs	+12.07%
⚡	Simulation	decode_primitives[u8, (1000, 2)]	44.6 µs	39.8 µs	+12.04%
⚡	Simulation	take[small_m/shuffled/primitive/nonnull/chunks=2048/indices=16]	376.2 µs	338.5 µs	+11.16%
🆕	WallTime	dict_canonicalize_gt_u8_neon[1000000]	N/A	50.4 µs	N/A
🆕	WallTime	dict_canonicalize_gt_u8_neon[16000000]	N/A	778.1 µs	N/A
🆕	WallTime	dict_canonicalize_gt_u8_avx2[1000000]	N/A	56.1 µs	N/A
🆕	WallTime	dict_canonicalize_gt_u8_avx2[16000000]	N/A	1.4 ms	N/A
🆕	WallTime	dict_canonicalize_gt_u8_avx512[1000000]	N/A	53.5 µs	N/A
🆕	WallTime	dict_canonicalize_gt_u8_avx512[16000000]	N/A	1.3 ms	N/A

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ji/small-u8-table-take-avx2 (0129184) with develop (62648ef)²}

Footnotes

1. 54 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on ji/small-u8-table-take-neon (21fa748) during the generation of this report, so develop (62648ef) was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩