Speed up batch scans and Python calls

[mrecachinas]

Why

For small inputs, the Hamming-distance calculation itself is already extremely fast. Most of the remaining time was being spent around it: setting up Python buffers, checking the selected algorithm repeatedly, or creating too many parallel tasks for batch scans.

This PR removes that avoidable work while preserving the existing API and results.

What changed

• Batch searches do less work. A best search stops as soon as it finds an exact match, and large scans use four coarse Rayon jobs instead of oversubscribing every CPU.
• Python calls take a shorter path. Exact bytes inputs bypass the general buffer protocol, and small generic buffers avoid an unnecessary GIL release/reacquire cycle.
• ARM hot paths exit earlier. Hex threshold checks inspect the first useful blocks sooner, and AArch64 uses LLVM’s faster native byte loop by default.
• Benchmarks and documentation are current. The README records the measured M4 Max results and explains where Python overhead still dominates.

Representative results

Measured on an Apple M4 Max using three independent runs:

Workload	Before	After	Improvement
Rust 512×16 best, exact match first	1,594 ns	8.2 ns	99.5%
Rust 16,384×64 best	213.3 µs	67.2 µs	68.5%
Rust 16,384×64 all	205.5 µs	76.9 µs	62.6%
Rust 100,000×128 best	330.1 µs	49.6 µs	85.0%
Python 64-byte distance	105.4 ns	46.6 ns	55.8%
Python 512×16 best, exact match first	1,958 ns	75.4 ns	96.1%

The single-record Rust kernels are now close to their practical hardware floors. PR #52 builds on this branch to address the remaining measured overhead in array dispatch, generic Python buffers, and concurrency thresholds.

Safety and compatibility

• No public API changes.
• Array ordering and lowest-index tie behavior are preserved.
• Large Python operations still release the GIL so other threads can run.
• Unsupported SIMD paths continue to use the existing fallbacks.

Testing

• cargo test --no-default-features
• cargo fmt --check
• ruff check .
• ruff format --check .
• python -m pytest -vls .
• Three Criterion benchmark passes
• Three pytest-benchmark passes

[github-actions]

📊 Benchmark Comparison Results

Status	Benchmark	Base	PR	Delta
⚠️	check_bytes_arrays_first_within_dist[16384 elems,s=64,end]	0.1288ms	0.1440ms	11.8% slower
⚠️	check_bytes_arrays_first_within_dist[16384 elems,s=64,mid]	0.0646ms	0.0718ms	11.2% slower
✅	check_bytes_arrays_best_within_dist[ 512 elems,s=16,at 0]	0.0053ms	0.0001ms	97.6% faster
✅	check_bytes_arrays_best_within_dist[ 512 elems,s=16,mid]	0.0050ms	0.0024ms	51.5% faster

➖ 41 benchmarks within noise (click to expand)

Benchmark	Base	PR	Delta
check_bytes_arrays_all_within_dist[ 512 elems,s=16,at 0]	0.0039ms	0.0042ms	9.3% slower
check_bytes_arrays_all_within_dist[ 512 elems,s=16,end]	0.0039ms	0.0043ms	9.2% slower
check_bytes_arrays_all_within_dist[ 512 elems,s=16,mid]	0.0039ms	0.0043ms	9.5% slower
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,at 0]	0.0083ms	0.0092ms	10.1% slower
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,end]	0.0083ms	0.0092ms	9.9% slower
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,mid]	0.0084ms	0.0092ms	9.8% slower
check_bytes_arrays_all_within_dist[16384 elems,s=64,at 0]	0.0878ms	0.0785ms	10.5% faster
check_bytes_arrays_all_within_dist[16384 elems,s=64,end]	0.0889ms	0.0786ms	11.6% faster
check_bytes_arrays_all_within_dist[16384 elems,s=64,mid]	0.0877ms	0.0782ms	10.9% faster
check_bytes_arrays_best_within_dist[ 512 elems,s=16,end]	0.0046ms	0.0048ms	3.2% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,at 0]	0.0098ms	0.0100ms	1.7% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,end]	0.0099ms	0.0100ms	1.8% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,mid]	0.0099ms	0.0100ms	1.6% slower
check_bytes_arrays_best_within_dist[16384 elems,s=64,at 0]	0.0957ms	0.0864ms	9.8% faster
check_bytes_arrays_best_within_dist[16384 elems,s=64,end]	0.0946ms	0.0858ms	9.3% faster
check_bytes_arrays_best_within_dist[16384 elems,s=64,mid]	0.0953ms	0.0859ms	9.9% faster
check_bytes_arrays_first_within_dist[ 512 elems,s=16,at 0]	0.0002ms	0.0001ms	54.5% faster
check_bytes_arrays_first_within_dist[ 512 elems,s=16,end]	0.0041ms	0.0045ms	8.2% slower
check_bytes_arrays_first_within_dist[ 512 elems,s=16,mid]	0.0022ms	0.0023ms	2.1% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,at 0]	0.0002ms	0.0001ms	55.1% faster
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,end]	0.0089ms	0.0097ms	9.2% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,mid]	0.0046ms	0.0049ms	7.2% slower
check_bytes_arrays_first_within_dist[16384 elems,s=64,at 0]	0.0002ms	0.0002ms	0.8% slower
check_bytes_within_dist[127 bytes,d=500]	0.0002ms	0.0001ms	53.8% faster
check_bytes_within_dist[16 bytes,d=0]	0.0002ms	0.0001ms	54.1% faster
check_bytes_within_dist[64 bytes,d=100]	0.0002ms	0.0001ms	58.0% faster
hex_within_dist	0.0001ms	0.0001ms	6.0% faster
hd_bytes[1000-diff]	0.0002ms	0.0001ms	43.2% faster
hd_bytes[1000-same]	0.0002ms	0.0001ms	37.7% faster
hd_bytes[1024-diff]	0.0002ms	0.0001ms	44.2% faster
hd_bytes[1024-same]	0.0002ms	0.0001ms	41.7% faster
hd_bytes[3-diff]	0.0001ms	0.0001ms	46.3% faster
hd_bytes[3-same]	0.0001ms	0.0001ms	43.3% faster
hd_bytes[64-diff]	0.0001ms	0.0001ms	43.3% faster
hd_string[1000-diff]	0.0003ms	0.0002ms	35.2% faster
hd_string[1000-same]	0.0002ms	0.0002ms	1.0% slower
hd_string[1024-diff]	0.0002ms	0.0002ms	2.0% slower
hd_string[1024-same]	0.0002ms	0.0002ms	13.7% slower
hd_string[3-diff]	0.0002ms	0.0002ms	no change
hd_string[3-same]	0.0001ms	0.0001ms	5.1% faster
hd_string[64-diff]	0.0001ms	0.0001ms	3.6% slower

Legend: ✅ Faster (>5%) · ⚠️ Slower (5-30%) · ❌ Regression (>30%) · 🆕 New · 🗑️ Removed