add ncclCommMarkAbort api in convienence of upper ai system controller

[woodlgz]

as issue 279 and issue 549 discussed, it's up to upper ai system controller to handle error when some rank losts.

it 's common for upper ai system controller to call ncclCommAbort to abort communicator related operations when rank lost situation is detected.
however, in system where multiple nccl communicators get involved, say horovod with HOROVOD_NUM_NCCL_STREAMS set to above 1, when lost rank is detected, trying to abort or destroy all these communicators will cause system hang. It's due to cuda implicit synchronization.
Image a scenario with 2 communicator (say A,B), system tries to abort all nccl communicators in sequence of A,B. If cuda kernel related to communicator B keeps running and won't exit until abort is detected while controller try to free communicator A(cudaFree involved), deadlock is met.

mark all communicator as aborted and give up all operations first, and release related resources later can avoid above problem.

[woodlgz]

@alsrgv take a look at this please