perf: add direct subgroup scan path

[SamJSui]

Summary

• remove the full input-to-output copy from the top scan pass
• use a feature-gated, coalesced one-item-per-lane subgroup scan
• retain and directly test the portable no-subgroup fallback
• publish process-isolated RTX and Jetson comparisons against pinned Massively 0.96

Why it was slower

The corrected scan still copied the entire input before doing useful work. At 100M items that was a 400 MB GPU copy. Its VT=8 shader also made neighboring lanes access values 32 bytes apart and used a Hillis-Steele workgroup scan with 16 barriers. Massively reads separate input/output buffers, accesses contiguous lane data, and uses subgroup operations with two workgroup barriers.

Results

Speedup is Massively time divided by wgpu-primitives time.

System	Scan 10M	Scan 100M	Compact 10M	Compact 100M
RTX 4070 Ti SUPER	2.81x	1.09x	1.65x	1.11x
dopey, 8-TPC Orin	1.53x	1.19x	1.28x	1.10x
grumpy, 4-TPC Orin	1.49x	1.18x	1.26x	1.09x

Jetson 10M bounded and full-width sort controls stayed within 0.12% of their published baselines.

Full medians and method: benchmarks/2026-08-08-direct-subgroup-scan.md.

Validation

• cargo fmt --all --check
• cargo test --lib --tests (63 passed)
• dedicated no-SUBGROUP fallback test
• 4,194,305-item three-level hierarchy regression
• cargo clippy --all-targets --all-features -- -D warnings
• cargo doc --no-deps
• cargo package --allow-dirty
• 10M and 100M correctness readback in every benchmark process on RTX, dopey, and grumpy
• Jetson power/clocks restored and live dynamic ranges verified after measurement