refactor: remove parallel GPU coordinator, fix streaming backpressure

[ChrisLundquist]

Summary

• Remove GPU coordinator from parallel scheduler (~900 lines deleted) — it serialized entropy encoding on one thread, bottlenecking at 28 MiB/s. The parallel path is now CPU-only (6+ GiB/s).
• Fix one-way backpressure ratchet in streaming GPU path — the coordinator now decrements pressure by batch_len on batch completion. Previously, once pressure exceeded the limit no try_send was attempted, so pressure could never decrease, permanently locking out GPU. GPU routing went from ~16% → ~62% on the mozilla corpus.
• Honor Backend::WebGpu in compress_with_options — multi-block GPU requests now route through compress_stream internally (in-memory Cursor I/O), which has the GPU coordinator with adaptive backpressure. Output uses framed format, transparently decompressible by decompress().

Additional cleanup

• Flatten UnifiedTask from single-variant enum to type alias (usize, usize)
• Simplify complete_task_lifecycle (remove dead next_task parameter, always None)
• Remove dead GPU telemetry methods and atomic fields from LocalSchedulerStats
• Add criterion benchmark (gpu_parallel_vs_streaming) comparing all four paths
• Update docs: CLAUDE.md, DESIGN.md, gpu-strategy.md, pipeline-architecture.md

Architecture after this PR

Path	GPU handling
Single block (input ≤ block_size)	compress_block respects Backend::WebGpu directly
Single thread (threads = 1)	Sequential compress_block respects Backend::WebGpu directly
Multi-thread + CPU	compress_parallel — CPU-only unified scheduler
Multi-thread + GPU	Routes through compress_stream — GPU coordinator with adaptive backpressure

Test plan

• All 694 tests pass (692 unit + 2 doc), clippy clean
• Pre-commit hook passes (fmt, clippy, build, test)
• Verified backpressure fix with instrumented streaming runs on Silesia corpus
• Criterion benchmark confirms parallel GPU path no longer bottlenecks (was 568ms → now 2.97ms, uses CPU-only path)
• Verify GPU streaming throughput on target hardware with cargo bench -- gpu_par_vs_stream

🤖 Generated with Claude Code