[TRAIN-2] DTensor async GRPO with dis-colocated generation hangs when tensor_parallel=2

## Problem

Dis-colocated async GRPO on the DTensor backend hangs during training when `tensor_parallel` is set to 2. Setting `tensor_parallel=1` unblocks training. This points to a synchronization issue that is specific to TP > 1.

## Steps to reproduce

1. Configure dis-colocated async GRPO with the DTensor backend and `tensor_parallel: 2`.
2. Start training and observe a hang with no error output.
3. Set `tensor_parallel: 1` and confirm training proceeds normally.

## Suggested fix

Audit synchronization barriers in the DTensor async rollout path for behavior that differs by tensor parallelism count.