Accelerate SHA-256 hashing on aarch64

[zaniebot]

uv#19662 started hashing every downloaded wheel and source distribution so that a cached distribution can be reused when a later operation requires its hash. On ARM, sha2 computes those hashes in software.

Enable the sha2 assembly backend on non-Windows aarch64, allowing wheel and source distribution downloads to use ARM SHA-256 instructions. This preserves the caching behavior from uv#19662 and also speeds up uv sync, which already verified lockfile hashes before that change.

I ran uv's standard cold and warm installation benchmarks against Trio's 41 documentation dependencies on a dedicated Apple Silicon Namespace runner. Each version was measured in 40 randomized, interleaved rounds. Times are means; changes are relative to main with paired-bootstrap 95% confidence intervals from 20,000 resamples.

Wall time (ms):

Benchmark	0.11.18	0.11.19	main	This change	Change (95% CI)
Cold uv pip sync	602.8	689.8	709.5	646.0	-8.9% [-11.1%, -6.8%]
Cold uv sync	533.2	500.1	504.4	455.1	-9.8% [-11.5%, -8.0%]
Warm uv pip sync	264.7	263.5	270.2	258.9	-4.2% [-10.2%, +2.1%]
Warm uv sync	48.6	49.8	52.2	51.8	-0.7% [-5.0%, +3.3%]

User CPU time (ms):

Benchmark	0.11.18	0.11.19	main	This change	Change (95% CI)
Cold uv pip sync	232.9	300.4	299.5	245.9	-17.9% [-19.0%, -16.9%]
Cold uv sync	292.1	291.3	292.1	241.6	-17.3% [-18.3%, -16.3%]
Warm uv pip sync	15.8	15.4	16.4	15.9	-2.9% [-7.2%, +1.5%]
Warm uv sync	9.7	9.9	11.1	10.9	-2.0% [-4.4%, +0.4%]

0.11.18 predates uv#19662, and 0.11.19 includes it. That change increased cold package-install wall time by 14.4% [11.3%, 17.9%] and user CPU by 29.0% [26.5%, 32.2%]. Hardware acceleration returns package-install user CPU to within 5.6% of the original baseline while preserving distribution hashes. Warm-install intervals overlap zero.

Separately, hashing a 9.7 MiB wheel 40 times improved from 1,076.5 ms to 208.2 ms on macOS ARM (5.2×) and from 1,234.2 ms to 204.2 ms on Linux ARM (6.0×).

[astral-automations-bot]

Promoted to #20805.