fix[gpu]: retain device buffers for dyn dispatch kernel

[0ax1]

fixes: #7838 (comment)

[joseph-isaacs]

Can we use a lifetime to catch this going forwards?

[codspeed-hq]

Merging this PR will not alter performance

⚠️ 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.

⚡ 1 improved benchmark
❌ 1 regressed benchmark
✅ 1219 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	chunked_varbinview_opt_into_canonical[(1000, 10)]	239.9 µs	202.5 µs	+18.49%
❌	Simulation	chunked_varbinview_opt_canonical_into[(1000, 10)]	188 µs	225.2 µs	-16.51%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ad/retain-buffers-for-launch (b0e0f2a) with develop (c8d915a)}

[0ax1]

Can we use a lifetime to catch this going forwards?

Dynamic dispatch is really a special case, as we use the raw u64 pointer values, and not CUDA views like for other kernels. For all other kernels, CUDA views are passed to the launch args. cudarc under the hood takes care of retaining read records like we do here for dyn dispatch manually.

In general, I'm not sure if this is solvable with lifetimes given that we describe buffer ownership across host and device, which can be freed C async style.

Maybe there's a way to more fundamentally rework our CUDA setup with: https://github.com/NVlabs/cuda-oxide or get inspiration from that.