feat(io): detect input compression by magic bytes, optional bz2/zstd/xz

[BenjaminDEMAILLE]

Closes #218.

• Input format is detected from the first bytes (files, pipes, stdin), not the extension. --readFilesCommand still takes precedence.
• gzip is always decoded with MultiGzDecoder (multi-member input is read in full).
• Optional pure-Rust features: bz2 (bzip2 0.6), zstd (ruzstd 0.9, with a multi-frame adapter), xz (lzma-rust2 0.21), and compressed-input for all three. Default features are unchanged.
• niffler was not used: it only offers C backends for zstd and xz.
• A detected format whose feature isn't built stops with an error naming the feature and a --readFilesCommand alternative.

Behaviour change to review: a plain file named *.gz is now read as plain text, and gzip is decoded whatever the filename. Documented in DIVERGENCE.md §4.3. The Parameters_openReadsFiles.cpp citation there should be double-checked.

Conflicts expected: #219 (same MultiGzDecoder change in fastq.rs), #225 (rapidgzip would move into the Gzip arm of compression.rs; both add [features]).

Tests: fmt, clippy 0 warnings, cargo test with default and all features. Test data is generated at runtime.

🤖 Generated with Claude Code