[libkineto] Refactor 4/n: Simplify activity logger step 2/3 (#68329)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68329
Pull Request resolved: https://github.com/pytorch/kineto/pull/466
1. Generalize ChromeTraceLogger::handleGenericActivity to enable it to handle Cuda runtime activities as well as the Roctracer generic activities.
This primarily involves enabling generic support for CPU -> GPU flows.
2. In the event of out-of-order GPU activities (an issue with Cuda11.0, likely fixed in later versions), no longer remove them but print warnings. Another diff will add these warnings to the metadata section.
Reviewed By: briancoutinho
Differential Revision: D31624496
fbshipit-source-id: dab04b3e3c0dd6799496ac87f837363de79eea25