cuda-checkpoint hangs when locking multiprocess nccl-tests

[tkj666]

driver version: 575.57.08
task: nccl-tests launched with openmpi

Process 1 can be checkpointed smoothly, but when process 1 is in locked/checkpointed state, executing cuda-checkpoint on process 2 will hang (any action including --get-state) until process 1 is unlocked. When process 1 is unlocked, the action on process 2 will be carried out, but if process 2 is in locked/checkpointed state, process 1 is also unable to interact with cuda-checkpoint.

[rst0git]

@tkj666 This is a known limitation - checkpointing applications that use NCCL is not supported yet.