feat[cuda]: add a cuda stream pool

[joseph-isaacs]

Hopefully it will help not to OOM gpus

[codspeed-hq]

Merging this PR will degrade performance by 29.5%

⚠️ 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.

⚡ 8 improved benchmarks
❌ 4 regressed benchmarks
✅ 1167 untouched benchmarks
⏩ 1323 skipped benchmarks¹

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	1M_10pct[100000]	21.4 µs	27.1 µs	-21.14%
⚡	Simulation	canonical_into_non_nullable[(10000, 1, 0.0)]	27.9 µs	24.7 µs	+12.88%
⚡	Simulation	canonical_into_non_nullable[(10000, 1, 0.01)]	34.2 µs	31.1 µs	+10.02%
⚡	Simulation	canonical_into_non_nullable[(10000, 10, 0.0)]	285.3 µs	194.1 µs	+46.95%
⚡	Simulation	canonical_into_non_nullable[(10000, 10, 0.1)]	471.6 µs	380.6 µs	+23.91%
❌	Simulation	canonical_into_non_nullable[(10000, 100, 0.01)]	2.2 ms	2.9 ms	-26.86%
❌	Simulation	canonical_into_non_nullable[(10000, 100, 0.1)]	3.7 ms	4.6 ms	-18.18%
⚡	Simulation	into_canonical_non_nullable[(10000, 10, 0.1)]	461.8 µs	383.9 µs	+20.3%
⚡	Simulation	canonical_into_non_nullable[(10000, 10, 0.01)]	312.1 µs	221.1 µs	+41.2%
⚡	Simulation	into_canonical_non_nullable[(10000, 10, 0.01)]	302.8 µs	228.1 µs	+32.78%
⚡	Simulation	into_canonical_non_nullable[(10000, 10, 0.0)]	276.1 µs	200.6 µs	+37.67%
❌	Simulation	canonical_into_non_nullable[(10000, 100, 0.0)]	1.9 ms	2.7 ms	-29.5%

---

_{Comparing ji/fuzz-oom (2a4cc78) with develop (f568de5)}

Footnotes

1. 1323 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. ↩

[0ax1]

fetch_add

[0ax1]

make sure capacity is 2 ^ n. and do % len on the returned value.

[0ax1]

good