[FEA] Record GPU time and Fetch time separately, instead of recording Total Time

Is your feature request related to a problem? Please describe. The totalTime metric can be confusing because it often (but not always) includes the execution time of the child operator(s). It would be better to separately record fetchTime and gpuOpTime, although this may be difficult or not possible in some cases.

Making this change will also provide better metrics that we can use when profiling queries or developing a better cost model for CBO.

Describe the solution you'd like

Review the following operators and determine whether we can improve the metrics to record the GPU time and fetch time separately. File new issues for those that we plan on doing this for.

[x] GpuBroadcastExchangeExec - No action required
[x] GpuBroadcastHashJoinExec - No action required
[x] GpuBroadcastNestedLoopJoinExec - No action required
[x] GpuBroadcastToCpuExec - No action required
[x] GpuCartesianProductExec - No action required
[x] GpuCoalesceBatchesExec - No action required
[x] GpuHashAggregateExec - No action required
[x] GpuRowToColumnarExec https://github.com/NVIDIA/spark-rapids/issues/2820
[x] GpuShuffleCoalesceExec - No action required
[x] GpuShuffledHashJoinExec - No action required
[x] GpuSortExec - No action required
[x] GpuTopN - No action required
[x] HostColumnarToGpuExec - No action required
[x] GpuUnionExec #2586
[x] GpuExpandExec #2512
[x] GpuGenerateExec #2512
[x] GpuRangeExec #2512
[x] GpuColumnarToRowExec #2269
[x] GpuFilterExec #2461
[x] GpuGlobalLimitExec #2461
[x] GpuLocalLimitExec #2461
[x] GpuProjectExec #2461
[x] GpuWindowExec #2461

Describe alternatives you've considered None

Additional context None

NVIDIA / spark-rapids

[FEA] Record GPU time and Fetch time separately, instead of recording Total Time #2334