CORE: fix timeout handle for triggered post

[Sergei-Lebedev]

What

Correctly handle time out error for triggered task

• For TL NCCL need to use commAbort, otherwise commDestroy could hang
• For other TLs need to destroy executor during finalize

[Sergei-Lebedev]

cc @Fuzzkatt

[Fuzzkatt]

Fixes the following unit tests on upstream Pytorch CI:

• test_verify_model_across_rank_with_logger: https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7722
• test_verify_model_across_rank_without_logger: https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7729
• test_ddp_model_diff_shape_across_ranks: https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7753
• test_ddp_model_diff_num_params_across_ranks: https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7776

Sample command to reproduce: BACKEND="ucc" WORLD_SIZE=3 python test/distributed/test_distributed_spawn.py -v -k test_verify_model_across_rank_with_logger

In general, when passing in mismatched inputs across ranks, UCC would originally fail to catch the error within the function itself and propagate the error into a barrier timeout.

For example, when passing in the mismatched inputs on https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7690, it would propagate the error into https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7707 instead of logging it properly within https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/distributed/distributed_test.py#L7695.

With this fix, UCC catches the error within the function correctly.