[FEA] Add companion metrics for all nsTiming metrics to measure time elapsed excluding semaphore wait

[binmahone]

Sometimes a huge number is observed in metrics, like:

This may confuse performance tuners because such metrics grew huge only because the task spent a lot of time waiting for the GPU semaphore. So for each timing metrics, like "concat time" or "op time", it is necessary to distinguish between w and w/o semaphore wait time:

If w is huge, and w/o is minor, then we should focus on issues like GPU contention.
If both w and w/o is huge, then we're more confident that it's more of a kernel/algorithm issues in single task.