Create the continuous batching CPU group with local synchronization (#48302)

* Create the continuous batching CPU group with local synchronization

DistributedHelper creates the TP CPU (gloo) group with a plain new_group, which requires every
rank of the default group to participate in the call. When the TP group is a subset of the world
(one continuous batching manager per node, data parallelism across nodes), the concurrent
per-subset calls deadlock in gloo's connectFullMesh until timeout. use_local_synchronization
restricts the creation to the group members, which is exactly the situation here; for the
world-spanning TP case it is a no-op behavior-wise.

* Trim comment