ENH: GPU idle_time and assorted GPU fixes

[lithomas1]

• Account for idle time on GPU (calculated by taking total time and subtracting time in events)
• Add parent/child relationships for annotations (restores hierarchical nature of NVTX ranges)
• Refactor exclusive metrics calculation to account for overlap between parent/child event (since child event can be on a different device and be launched after parent event finishes)

[Copilot]

Pull Request Overview

This PR enhances GPU trace analysis by accounting for GPU idle time and fixes issues related to GPU event annotations and exclusive metric calculations.

• Refactors the caller/callee matching logic using groupby and a helper function.
• Updates exclusive metric calculations to correctly handle overlapping GPU event timings.
• Revises GPU idle time computation using groupby to streamline the calculation.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
pipit/trace.py	Refactored _match_caller_callee logic, updated exclusive metric calculation and idle time computation for improved handling of GPU overlaps and asynchronous execution.
pipit/readers/nsight_sqlite_reader.py	Updated DataFrame conversion and sorting, and refined parallelism level assignments based on trace types.

[jhdavis8]

Looks good, thanks @lithomas1 !

[lithomas1]

Thanks for the review!