docs: audit SIMD/scalar parity coverage across checksums crate

[oferchen]

Summary

• Document the runtime CPU-feature dispatch ladder for every SIMD-accelerated checksum (rolling, MD4, MD5, SHA-256, CRC32C) directly in the relevant module rustdoc, so the supported AVX-512 -> AVX2 -> SSE4.1 -> SSSE3 -> SSE2 -> NEON -> scalar fallback chain is discoverable without grepping for is_x86_feature_detected!.
• Add streaming-vs-one-shot parity tests for SHA-256 and CRC32C (16 KiB deterministic pseudo-random buffer + chunk-size sweep, plus RFC 3720 canonical vectors fed through the streaming wrapper for CRC32C) to guard against regressions in the streaming state machine that the existing one-shot vectors do not catch.
• No runtime behaviour changes - purely additive rustdoc and tests.

Per-algorithm gap analysis

Algorithm	Dispatch tiers documented	Existing parity tests	Gaps closed by this PR
Rolling	AVX2 -> SSE2 -> NEON -> scalar	rolling::tests::checksum::simd (lane sweeps, proptest, large inputs)	Module rustdoc with dispatch table only - no test gap
MD4	AVX-512 -> AVX2 -> SSE2 -> NEON -> WASM -> scalar	simd_parity_tests::md4_simd_parity (RFC 1320 + lane boundaries + proptest + 100 KiB inputs)	Dispatch ladder + SSSE3/SSE4.1 reuse note only - no test gap
MD5	AVX-512 -> AVX2 -> SSE4.1 -> SSSE3 -> SSE2 -> NEON -> WASM -> scalar	simd_parity_tests::md5_simd_parity (RFC 1321 + lane boundaries + proptest + 100 KiB inputs)	Dispatch ladder only - no test gap
SHA-256	SHA-NI -> ARMv8 sha2 -> sha2-asm -> pure-Rust	One-shot vectors + 1 KiB / 8 KiB streaming checks	Added 16 KiB streaming/one-shot parity + chunk-size sweep
CRC32C	x86_64 SSE4.2 -> aarch64 CRC -> software fallback	Canonical one-shot vectors + 4 MiB streaming throughput	Added 16 KiB streaming/one-shot parity + chunk-size sweep + canonical vectors via streaming API

Test plan

• CI fmt + clippy
• CI nextest (stable / Linux musl / macOS / Windows) - new tests run on every architecture and exercise whichever SIMD backend the host selects
• CI interop validation