[STF] Do not keep track of dangling events in a CUDA graph backend

[STF] Do not keep track of dangling events in a CUDA graph backend #4922