perf(io): encode SAM/BAM records on the align workers (stacked on #276)

[BenjaminDEMAILLE]

Stacked on #276 (base perf/parallel-bgzf-writer). Refs #223: the follow-up that moves record encoding off the single writer thread.

• New src/io/encode.rs: RecordEncoder produces SAM/BAM bytes through the same noodles writers the output uses (NoQS stripping, CIGAR check included).
• Each batch's records (main, WithinBAM supplementary, TranscriptomeSAM) are encoded in parallel on the align workers; writers gain write_encoded.
• Sorted BAM: SortBuffer stores encoded records and stable-sorts an index on (ref id, start), unmapped last, same order as before.
• Gated: the old writer thread is effectively a free extra core, so pre-encoding costs 0.5 to 1.3s at 1 to 8 threads. It turns on when runThreadN + spare > available cores (spare 4 for BAM, 6 for SAM). BySJout always keeps the old path. RUSTAR_PRE_ENCODE=1|0 forces it.

Wall time (s), 16-core Mac, 2M SE 100 bp reads, median of 3:

threads	None	BAM main	BAM #276	BAM this PR (auto)	SAM main	SAM forced on
1	12.55	15.19	13.79	14.04	14.11	14.37
8	2.87	3.07	2.99	3.00	3.03	3.19
12	2.33	3.36	2.51	2.52	2.87	2.48
16	2.26	3.94	2.69	2.42	3.19	2.41

Identity (cmp vs main, forced on): SAM, BAM unsorted/sorted, --outStd, NoQS, BySJout, TranscriptomeSAM, GeneCounts, unmapped Fastx, PairedKeepInputOrder, two-pass, PE chimeric WithinBAM, solo: all identical. tests/pre_encode_parity.rs covers 14 configurations on/off.

Conflicts: 3-way merge simulation of src/lib.rs against #222 and #261 gives 0 conflicts. Nearest hunks are the solo batch_size lines; if they conflict, keep both lines. #261 note: AlignmentBatchResults grows by 48 bytes, so batches get slightly smaller (no output change).

Pre-existing, not fixed here: SE --chimOutType WithinBAM fails on main ("read length-sequence length mismatch"); BufferedSamRecords::new() reserves 10,000 records per read.

Tests: fmt, clippy 0 warnings, cargo test 639 passed.

🤖 Generated with Claude Code