perf: CUDA async to host copy

[0ax1]

No description provided.

[codspeed-hq]

Merging this PR will degrade performance by 31.44%

⚡ 4 improved benchmarks
❌ 3 regressed benchmarks
✅ 1247 untouched benchmarks
⏩ 1254 skipped benchmarks¹

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

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	canonical_into_non_nullable[(10000, 10, 0.01)]	219.7 µs	308.6 µs	-28.82%
⚡	Simulation	canonical_into_non_nullable[(10000, 1, 0.01)]	44.4 µs	37.1 µs	+19.5%
⚡	Simulation	canonical_into_non_nullable[(10000, 1, 0.1)]	59.4 µs	53.2 µs	+11.61%
❌	Simulation	canonical_into_non_nullable[(10000, 10, 0.0)]	192.9 µs	281.4 µs	-31.44%
⚡	Simulation	canonical_into_non_nullable[(10000, 1, 0.0)]	38.9 µs	31.8 µs	+22.45%
❌	Simulation	canonical_into_non_nullable[(10000, 10, 0.1)]	375.4 µs	468 µs	-19.78%
⚡	Simulation	canonical_into_nullable[(10000, 100, 0.0)]	5.1 ms	4.1 ms	+24.85%

---

_{Comparing ad/async-copy (afbc61c) with develop (1a0d672)}

Footnotes

1. 1254 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. ↩