ProcessGroupNCCL,Manager: surface async abort errors correctly

[d4l3k]

This adds a mechanism for propagating abort errors to Manager via ProcessGroup.errored().

Without this change we don't wait for all operations to complete in Manager.should_commit and even if an abort occurs we don't detect it as they're running asynchronously on the GPU and NCCL abort causes it to return successfully.

This intentionally adds a synchronization point in should_commit.

This also includes a small fix in the resiliency tests to avoid stream dependencies between different worker threads.

Test plan:

pytest -o log_cli=1 torchft/process_group_test.py -v -s -x -k 'NormalNcclMultiPgTest'
pytest torchft/manager_test.py