rapidgzip-rust v0.3.0

Added

• Opt-in DecoderPool process-wide decode budgets with a validated bon
  builder, runtime resizing, stable max-min cross-decoder allocation, fair
  contended admission, and aggregate queue/thread/activity telemetry. Existing
  decoders retain private elastic scheduling unless a pool is attached.
• Persistent runtime growth through DecoderHandle::request_workers and
  reader forwarding methods. DecoderStats now exposes the explicit request,
  desired concurrency before pool contention, and whether the shared pool is
  limiting progress.
• A concurrent programmatic-reader benchmark driver for shared versus private
  scheduling through both ordinary Read and paraseq FASTQ parsing.
• Optional cpu-accounting feature reporting lifetime worker and auxiliary
  thread CPU through DecoderStats::completed_worker_cpu_time,
  completed_auxiliary_cpu_time, and cpu_accounting_failures. Each thread
  reads its own CPU clock exactly twice, at registration and exit, so no timing
  call, atomic, or branch is added to decode task begin/end or the inflate
  loops. With the feature disabled the accounting compiles out and the fields
  are None. Threads still running are omitted until exit, making this
  component accounting rather than an exact decision-window counter.
• Optional busy-time-accounting feature exposing a monotonic cumulative
  executing-region integral as DecoderStats::accounted_busy_time. This gives
  a scheduler an exact busy-time signal that does not depend on its own polling
  cadence; reconstructing the same quantity by sampling executing-worker counts
  was measured 17-313% off on sparse, bursty, and stored decoder paths.

Changed

• Marker/window admission now uses immutable configured decoder and pool
  capacity rather than the transient application ceiling. A low early throttle
  bounds screen execution but no longer permanently latches a suitable stream
  to the sequential path, allowing reliable later growth.
• Pool execution accounting covers specialized workers, sequential chunks,
  bounded format scans, coordinator bridges/fallbacks, and ordered checksums.
  Permits are released before bounded result and final reader handoffs so a
  backpressured decoder cannot strand capacity needed by another file.
• Shared BGZF and dense-member workers amortize global permit traffic with a
  worker-local lease that is retained only across nonblocking handoffs and
  immediately available work. Shared positional readers sample an
  allowance-sized logical output backlog only when the pool is divided, so
  broad per-file headroom does not multiply persistent parser buffering or
  penalize a sole fast reader.