avoid cyclic ownership with GPU timers

[upsj]

Discovered by @MarcelKoch, ProfilerHook::create(_nested)_summary doesn't work with GPU timers because they keep the executor alive via a cyclic dependency. This breaks up the dependency by storing device ID and stream explicitly.

Previously, attaching a ProfilerHook logger with a GPU timer lead to a dependency graph like this:

graph TD;
Executor --> ProfilerHook;
ProfilerHook --> CudaTimer;
CudaTimer -->|stream, device_id| Executor;

Loading

The new dependency graph removes the cycle

graph TD;
Executor --> ProfilerHook;
ProfilerHook --> CudaTimer;
CudaTimer --> stream;
CudaTimer --> device_id;

Loading

[pratikvn]

Cool graphs! Didnt know GFM was capable of that. LGTM!

[sonarqubecloud]

Quality Gate passed

The SonarCloud Quality Gate passed, but some issues were introduced.

5 New issues
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarCloud