GPU: keep patch indices on device

[joseph-isaacs]

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

What

• Preserve patch-index buffers in device memory through fused dispatch.
• Consume them directly from CUDA kernels.
• Remove the stream synchronization that existed only to materialize indices on the host.

Why this is needed

Patched BitPacked/ALP-style encodings are normal compressor output. The old fused path converted patch indices back to a host primitive and synchronized the stream before the consumer kernel, even though both producer and consumer are ordered on the same stream. Keeping indices in their device buffer removes that D2H boundary and the forced synchronization while preserving the same patch semantics.

Validation

Checked independently against this PR's current base:

• cargo check -p vortex-cuda --all-features

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 14.03%

⚠️ 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.

❌ 7 regressed benchmarks
✅ 2036 untouched benchmarks
⏩ 46 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	cuda/alp_f64/10%[100M]	6.9 ms	8.4 ms	-17.79%
❌	WallTime	cuda/alp_f32/0%[100M]	2.5 ms	2.9 ms	-14.98%
❌	WallTime	cuda/alp_f64/1%[100M]	6.6 ms	7.7 ms	-14.36%
❌	WallTime	cuda/alp_f32/1%[100M]	4.8 ms	5.6 ms	-14.28%
❌	WallTime	cuda/alp_f32/10%[100M]	4.5 ms	5.1 ms	-13.01%
❌	WallTime	words_gather_scalar[65536]	8.2 µs	9.4 µs	-12.49%
❌	WallTime	words_gather_dispatch[1024]	8 ns	9 ns	-11.11%

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-device-patch-indices (67c9110) with develop (b825c4f)}

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. ↩