fix(api): bound native PDF extraction concurrency with a semaphore

Unbounded spawn_blocking extractions let many large PDFs pile onto one
pod's heap at once (each holds the <=50MB file plus parsed text/markdown),
contributing to OOM kills. Bound per-process concurrency via a small
FIFO semaphore (PDF_EXTRACTION_CONCURRENCY, default 3).

Expose Prometheus metrics to tune the limit:
- pdf_extraction_semaphore_active (gauge)
- pdf_extraction_semaphore_waiting (gauge)
- pdf_extraction_semaphore_wait_seconds (histogram)
- pdf_extraction_semaphore_hold_seconds (histogram)