[core][experimental] Allow custom NCCL group declaration in accelerated DAG

[stephanie-wang]

Description

Currently, accelerated DAGs automatically create a NCCL group of actors based on the annotated DAG nodes. However, sometimes user code will already create a NCCL group, or they will want more explicit control over which actors are in the NCCL group, especially if there are multiple groups. To support this, we should allow the user to pass in custom NCCL groups when creating the accelerated DAG.

Use case

No response

[anyscalesam]

@stephanie-wang how important is this; is this needed to enable pipeline parallel training? Or is this just a general ADAG devex improvement?

[stephanie-wang]

No, it should not be necessary. Just a UX improvement.

[rkooo567]

this is an important feature for vllm integration (and is explicitly asked)