fix(dump): drain pending data in FIFO writer on shutdown

[nick-youngblut]

Summary

Fixes a data-integrity race in the multi-threaded dump writer where pending bytes could be silently dropped when the buffered FIFO writer shut down, truncating FASTA/FASTQ output (most visibly over named pipes).

Changes

• Replace ThreadWriter's dual signaling (an mpsc shutdown channel plus a Condvar over the buffer) with a single mutex-guarded WriterState { buffer, closed } driven by one condition predicate.
• The worker thread now exits only when the buffer is empty and closed is set; a non-empty buffer always takes the drain branch, so all ingested bytes are flushed before the thread exits.
• Drop sets closed = true under the lock before notifying, closing the wakeup gap, then joins the worker.
• ingest() adapted to the new state struct; backpressure behavior is unchanged.
• Add in-memory regression tests, including a 1000-iteration stress loop that ingests immediately before drop to surface the race.

[gemini-code-assist]

Code Review

This pull request refactors ThreadWriter to resolve a data-loss race condition during shutdown by introducing a shared WriterState struct that manages both the write buffer and the shutdown state under a single mutex and condition variable. Additionally, new unit tests are added to verify proper draining of pending data on drop. The review feedback highlights two key areas for improvement: first, a critical issue where a mutex lock is held during a sleep call in the ingest loop, which blocks the worker thread; and second, a performance optimization to reuse the write buffer's allocated capacity instead of discarding it on every write cycle.