feat: add accelerated IEEE CRC32 and CRC32C checksums

[tisonkun]

Summary

Add opt-in IEEE CRC32 and Castagnoli CRC32C for records, pages, and protocol payloads, with one-shot and incremental APIs, finalized-checksum resumption, segment combination, and std::io::Write integration. The library remains dependency-free, allocation-free, and no_std compatible. Raise the workspace MSRV from Rust 1.85 to 1.89 so the library and every benchmark comparison share one compiler baseline.

Accelerate both variants with AArch64 CRC/PMULL, fused CRC/SHA3 loops, and x86-64 PCLMULQDQ plus AVX2/AVX-512 VPCLMULQDQ. Keep a slicing-by-8 scalar fallback and differential tests for every available backend, unaligned inputs, streaming boundaries, and large shift lengths.

Design Notes

The API model groups general-purpose hashing, error-detecting checksums, compatible hashing, and legacy digests by their intended use.

Each CRC state stores one u32. digest preserves the state, and from_digest accepts a finalized checksum. The two polynomials have distinct names and types; these APIs consume bytes without Hasher adapters. Combine functions take the right segment's exact u64 length.

std enables runtime CPU detection; no_std uses compile-time target features. Folding schedules draw on corsix/fast-crc32 and crc-fast; their MIT notices are collected in LICENSE, which is included in the published crate. Kernel loads accept unaligned input and remain inside the slice.

The ecosystem benchmarks compare stock crc 3.4.0 (Table<1> and Table<16>), crc32fast 1.5.1, crc32c 0.6.8, and crc-fast 1.10.0. All comparisons are enabled by default, with no benchmark feature flags. Stable and Rust 1.89 CI both build every workspace target and run workspace tests. The suites cover short inputs, dispatch boundaries, one-byte offsets, and streaming chunks. CI compiles the benchmark targets without running performance measurements. Performance regressions are checked locally against the base revision with matching environments and sampling settings.

1 MiB one-shot medians measured at cd9e5ba below are Hashcrew / crc-fast, in microseconds, using Rust 1.98.0, std, default targets, and a reused buffer on one thread. Each value is the median of the four longer runs:

Host	CRC32C	IEEE CRC32
Linux, AMD EPYC 7763	41.19 / 55.99	41.17 / 82.35
macOS, Apple M1 runner	13.94 / 13.81	13.96 / 13.67
macOS, local Apple M4 Max	9.67 / 10.32	9.67 / 10.38

Bulk throughput is competitive across these hosts. The full M1 matrix still has roughly 6–13% gaps at some 16 KiB + 1 / 64 KiB cases and sub-nanosecond gaps at a few tiny inputs; this draft does not establish pointwise dominance. The shared M1 runner also shows substantial run-to-run variation. The AVX-512 path is compiled and has conditional differential coverage; these measurements exercise AVX2 and AArch64 hardware.

Measurements and CPU/toolchain details at cd9e5ba are preserved as artifacts. CI validation: Linux/macOS/Windows on Rust 1.89 and stable, feature/target checks, lint, and Miri. Native AArch64 Linux container tests and local macOS measurements also passed. The default-enabled benchmark suites passed a local cargo x bench -- --test smoke check, and package inspection verifies the consolidated LICENSE.