Remove PR/trunk distinction as it isn't accurate + fix win sharding
The PR/trunk distinction in CI has become confusing due to several changes and bugs:
1. parse_ref.py does not correctly output the branchname for windows workflows, so BRANCH is always empty for windows CI. This cause the print_test_stats script to upload the stats to the pr_test_time dir instead of test_time, which is where we look for the stats.
2. Ever since our migration to use push triggers for CIFlow PR workflows, many of them are triggered on push and PR_NUMBER, PR_LABELS, and PR_BODY are all unset, and any code relying on these variables is not guaranteed to be accurate. (https://github.com/pytorch/pytorch/issues/73022)
We could fix this by forcing our stats parsing code to look in both directories for reports, but that just adds confusion. We should take a step in fixing #73022 and start removing these distinctions.
Would close https://github.com/pytorch/pytorch/issues/72371
NOTE: after this change, all commit stats will go in the test_time directory.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73293