perf: add hex-simd benchmarks and optimize SIMD implementations

[zerosnacks]

Adds hex-simd to benchmarks (closes #25) and optimizes SIMD implementations to match/exceed hex-simd performance.

Benchmark results

CPU: Intel Core i9-13900K (x86_64, AVX2)

check

Size	const-hex (before)	const-hex (after)	hex-simd	faster-hex
32B	2.47 ns	2.44 ns	1.71 ns	3.36 ns
256B	17.58 ns	16.69 ns	13.50 ns	23.36 ns
2KB	150.6 ns	81.69 ns	113 ns	146.7 ns
16KB	1.13 µs	570.6 ns	885.6 ns	1.13 µs
128KB	8.96 µs	5.50 µs	6.99 µs	10.2 µs
1MB	68.73 µs	46.64 µs	53.29 µs	71.92 µs

encode

Size	const-hex (before)	const-hex (after)	hex-simd	faster-hex
32B	8.50 ns	8.98 ns	6.91 ns	12.03 ns
256B	23.69 ns	15.61 ns	16.22 ns	21.08 ns
2KB	88.69 ns	72.69 ns	107.3 ns	110.6 ns
16KB	599.6 ns	548.1 ns	845.6 ns	726.6 ns
128KB	4.42 µs	3.75 µs	6.64 µs	7.59 µs
1MB	47.43 µs	47.96 µs	53.70 µs	77.20 µs

encode_to_slice

Size	const-hex (before)	const-hex (after)	hex-simd	faster-hex
32B	1.90 ns	1.73 ns	1.71 ns	4.09 ns
256B	10.49 ns	8.28 ns	12.64 ns	13.38 ns
2KB	77.63 ns	60.10 ns	103.3 ns	85.44 ns
16KB	613.1 ns	539.9 ns	825.6 ns	653.9 ns
128KB	5.28 µs	5.11 µs	6.61 µs	5.22 µs
1MB	55.39 µs	55.40 µs	52.73 µs	51.51 µs

decode

Size	const-hex (before)	const-hex (after)	hex-simd	faster-hex
32B	28.69 ns	16.05 ns	9.46 ns	13.71 ns
256B	46.75 ns	26.16 ns	35.16 ns	43.66 ns
2KB	256.6 ns	161.8 ns	252.6 ns	270.4 ns
16KB	1.97 µs	1.28 µs	1.94 µs	2.03 µs
128KB	15.54 µs	9.62 µs	14.80 µs	15.98 µs
1MB	118.2 µs	82.16 µs	123.8 µs	135.8 µs

decode_to_slice

Size	const-hex (before)	const-hex (after)	hex-simd	faster-hex
32B	5.57 ns	7.20 ns	3.62 ns	6.49 ns
256B	30.77 ns	21.58 ns	28.66 ns	36.10 ns
2KB	249.3 ns	163.1 ns	235.1 ns	250.9 ns
16KB	1.96 µs	1.28 µs	1.97 µs	1.98 µs
128KB	15.55 µs	10.11 µs	14.85 µs	15.65 µs
1MB	118.2 µs	81.89 µs	124.5 µs	122 µs

format

Size	const-hex (before)	const-hex (after)	std
32B	13.95 ns	7.30 ns	549.6 ns
256B	20.35 ns	18.96 ns	3.81 µs
2KB	121.1 ns	108.5 ns	30.7 µs
16KB	1.22 µs	1.10 µs	135.5 µs
128KB	11.11 µs	11.04 µs	1.06 ms
1MB	151.1 µs	161.4 µs	8.45 ms

Changes

• Keep original 6-comparison SSE2 algorithm for small inputs (addresses, hashes)
• Add AVX2 check with signed overflow trick for larger inputs ≥128 bytes (Muła & Langdale)
• Double encode throughput (32→64 bytes/iter) via new AVX2 path
• Add 13 new edge case tests covering SIMD boundaries and all byte values

[DaniPopes]

Split into #38 (benchmarks) and #39 (optimizations). Thanks @zerosnacks!