GPU benchmark: add a single-stream execution fast path

[joseph-isaacs]

Independent of the other follow-up PRs; based directly on #9147. This draft contains one reviewable GPU correctness or performance concern.

What

• Execute independent field decodes directly in stream order for the benchmark.
• Retain the general scheduling path outside this explicit fast path.

Why this is needed

This is benchmark-specific and removes a different event class from #9436. The GPU benchmark executes all work on one CUDA stream, where FIFO ordering already prevents use-before-produce. cudarc's per-buffer events are needed for cross-stream ownership, not for this single-stream session. The profile observed 2,731 stream/event waits, so using a one-stream context with cross-stream tracking disabled removes pure bookkeeping while leaving the general multi-stream session unchanged.

Validation

Checked independently against this PR's current base:

• cargo check -p compress-bench --features cuda,unstable_encodings

The combined implementation also passed Rust/CUDA formatting, 164 focused dynamic-dispatch tests, and 28 focused constant-array tests.

[codspeed-hq]

Merging this PR will degrade performance by 1.74%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 2 improved benchmarks
❌ 2 regressed benchmarks
✅ 2033 untouched benchmarks
⏩ 46 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	compact[(4096, 90)]	1.6 µs	1.9 µs	-13.97%
❌	Simulation	compact[(16384, 90)]	1.8 µs	2 µs	-13.01%
⚡	Simulation	take[small_m/shuffled/primitive/nonnull/chunks=16384/indices=16]	1.2 ms	1 ms	+12.47%
⚡	Simulation	compress_fsst[(1000, 64, 8)]	1.1 ms	1 ms	+10.75%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing joe/gpu-single-stream-fast-path (599f86e) with claude/gpu-decompress-benchmarks-4mmn93 (8e060ca)}

Footnotes

1. 46 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩