perf[gpu]: reduce register pressure in dyn dispatch

[0ax1]

We decrease the number of values per tile in the output stage each GPU thread uses, as well as limit the register count to 32 in the launch bounds. This brings the dynamic dispatch kernel into a reasonably close range compared to the standalone kernel for now.

Type	Dynamic dispatch	Standalone	Ratio
u8 bw6	172 µs	79 µs	2.17×
u16 bw6	140 µs	88 µs	1.59×
u32 bw6	184 µs	148 µs	1.24×
u64 bw8	303 µs	276 µs	1.10×

[0ax1]

It's good to keep in mind, that dyn dispatch with only running bp is an odd scenario. We can spend more time on fine tuning bp in context of dyn dispatch but should prioritize optimizing end-to-end perf at this point.

[codspeed-hq]

Merging this PR will improve performance by 20.02%

⚡ 9 improved benchmarks
✅ 1154 untouched benchmarks
⏩ 1457 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	take_map[(0.1, 0.5)]	1,154.5 µs	980.3 µs	+17.77%
⚡	Simulation	take_map[(0.1, 1.0)]	2 ms	1.6 ms	+20.02%
⚡	Simulation	patched_take_10k_contiguous_patches	258.1 µs	227.7 µs	+13.32%
⚡	Simulation	patched_take_10k_dispersed	316 µs	285.8 µs	+10.58%
⚡	Simulation	patched_take_10k_contiguous_not_patches	258.4 µs	228.1 µs	+13.28%
⚡	Simulation	patched_take_10k_first_chunk_only	302 µs	271.8 µs	+11.14%
⚡	Simulation	take_10k_first_chunk_only	270.6 µs	225.7 µs	+19.89%
⚡	Simulation	patched_take_10k_random	270.3 µs	240 µs	+12.64%
⚡	Simulation	take_10k_dispersed	284.4 µs	239.5 µs	+18.76%

---

_{Comparing ad/cap-values-per-tile (ac24aaf) with develop (1169d84)}

Footnotes

1. 1457 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. ↩