[CUDA][HIP] Minimize native events recorded and created by urEnqueueTimestampRecordingExp

[hdelan]

trafficstars

Since EvStart and EvEnd are recorded directly after one another in urEnqueueTimestampRecordingExp, we can just copy EvStart to make EvEnd, instead of calling cuEventRecord for both EvStart and EvEnd, one after the other.

@steffenlarsen for discussion

[hdelan]

I now get what you meant! I'm alright with this solution. 👍

Great thanks. I'll add a DPC++ PR just to make sure no code is relying on EvEnd being after EvStart

[npmiller]

This ended up introducing some bugs and we have refactored the base event since, so closing this.