perf: parallel pread for SE/PE uncompressed FASTQ input

[KimBioInfoStudio]

Summary

• Parallel pread: Worker threads read uncompressed FASTQ input via pread(2) in parallel, bypassing the single reader thread bottleneck. A FastqChunkIndex scans the file once to build pack-aligned byte offsets, then each worker atomically grabs the next pack index, preads its chunk, and parses it into reads.
• Raw FASTQ pwrite: Extends the existing parallel pwrite(2) output (previously gz-only) to raw .fq output. Worker threads write directly to non-overlapping file regions, bypassing the single writer thread.
• Ordered output: WriterThread gains inputWithSeq(tid, data, seq) which uses pack index as sequence number — for pwrite mode via the offset ring, for non-pwrite mode (stdout) via a new ordered ring buffer drained by the writer thread.

Commits

1. perf: extend pwrite parallel write to raw FASTQ output — raw fq pwrite + hybrid spin backoff
2. perf: SE parallel pread with ordered output — SE parallel read + ordered writer mode
3. perf: PE parallel pread with ordered output — PE dual-file parallel read + all 7 writers ordered

When parallel pread activates

• Input must be uncompressed (not .gz)
• Input must be regular files (not stdin)
• Not in split mode
• No --reads_to_process limit
• PE: not interleaved input; R1/R2 must have same read count

Falls back to the existing sequential reader otherwise.

Benchmark (Apple M4 Pro, 10M PE reads, -w 14)

Mode	Baseline	Optimized	Speedup	Verify
se-fq-gz	14.04s	13.68s	1.03x	PASS
se-fq-fq	12.86s	12.88s	1.00x	PASS

Small -w shows no benefit on fast NVMe — the reader thread is not the bottleneck at low thread counts. Requesting benchmark on high-core-count x86 systems where the single reader thread may become a bottleneck with large -w.

@sfchen Could you benchmark this on a high-core x86 machine with -w 32 or higher on large uncompressed FASTQ files? The parallel pread should show more benefit when the sequential reader becomes the bottleneck.

Test plan

• SE fq→fq output matches baseline byte-for-byte
• SE fq→gz output matches baseline (decompressed content)
• SE fq→stdout output matches baseline
• PE fq→fq output matches baseline byte-for-byte (R1 + R2)
• Small file (10 reads, -w 8) completes without hang
• PE mode regression (sequential reader path unchanged)

🤖 Generated with Claude Code