Context: it's been a while I've been working on #8686 (_conj_copy lowering). The implementation itself was finished last week. However, I ran into a TPU CI failure, which I spent some time making it work by reverting my changes. Seeing that that worked, I tried bringing back the changes, which ended up working.

Problem: upon closer inspection, it seems that the CI always merges my PR to PyTorch/XLA master branch, before actually compiling (and testing) PyTorch/XLA. While testing it like that (merged PR) is useful for guaranteeing that we are not breaking anything, that seems to me like a source of flakiness that could be avoided.

Proposed Solution: test the PR as is on CI, by default

• Trigger the merged PR test:
  • After the PR was actually merged
  • By introducing a label to the PR, e.g.: ci/master