perf: parse GTF without per-line allocations (1.88x)

[BenjaminDEMAILLE]

What

Removes the per-line allocations from parse_gtf().

Before, each line cost a String (reader.lines()) plus a Vec<&str> for the split fields, the attribute column was scanned once per attribute wanted, and gene_id / transcript_id were cloned 3-4 times per line to build map keys. On a human annotation that is tens of millions of allocations.

• Read into a reused byte buffer with read_until instead of lines().
• Split the 9 tab-separated fields into a fixed [&str; 9] with memchr, stopping after the 9th field.
• Extract gene_id and transcript_id in one pass over the attribute column instead of two; get_attribute returns a borrowed &str so callers allocate only when keeping the value.
• Probe the gene map and the transcript-builder map with borrowed keys, allocating owned String keys only on insert. TxKeyRef implements Hash + indexmap::Equivalent so an IndexMap<(String, String), _> can be looked up with &str slices.

Measurement

parse_gtf() on Homo_sapiens.GRCh38.113.gtf (4,114,455 lines, 1.66 GB uncompressed), aarch64 macOS, hyperfine -w 1 -r 5:

input	before	after
plain .gtf	2.373 s ± 0.017 s	1.264 s ± 0.010 s	1.88x
gzipped .gtf.gz	2.997 s ± 0.015 s	1.924 s ± 0.034 s	1.56x

(The gzipped row is on top of the current miniz_oxide backend; it compounds with #138.)

Parity

Identical output on the same annotation before and after: 78,932 genes / 387,944 transcripts / 2,164,410 exons / 195,851,333 total effective length.

Behaviour deliberately preserved in the edge cases:

• lines with fewer than 9 fields are skipped, as split('\t').collect() + len() < 9 did;
• a 10th field is truncated at the tab, as split('\t') did;
• first occurrence wins when an attribute key repeats;
• CRLF line endings are handled like lines() did.

New unit tests cover each of those. cargo test --release passes (203 + 12 + 18 + 2).

---

🤖 Generated with Claude Code

Part of #143.