GB200 GRPO training hangs around step 3-4 during DTensor policy offload/refit

**Describe the bug**

GRPO training for DeepSeek V4 Flash hangs on GB200 around step 3-4. The job can finish the first few steps, then stops making progress during the next generation/refit cycle and eventually either needs to be cancelled or fails.

Debug log shows the hang is likely in the policy worker post-refit CPU offload path:

```text
DTensorPolicyWorkerV2[rank=5]: offload_after_refit/model_to_cpu/move_to_device_start ...
DTensorPolicyWorkerV2[rank=5]: offload_after_refit/model_to_cpu/move_to_device_done after 4106.03s ...

DTensorPolicyWorkerV2[rank=1]: offload_after_refit/model_to_cpu/move_to_device_start ...
DTensorPolicyWorkerV2[rank=1]: offload_after_refit/model_to_cpu/move_to_device_done after 5133.73s ...

[rank3] Watchdog caught collective operation timeout:
WorkNCCL(SeqNum=64214, OpType=ALLGATHER, NumelIn=1, NumelOut=32, Timeout(ms)=600000)
```

The same script previously ran on H100, but repeatedly hangs on GB200.

**Steps/Code to reproduce bug**

Run 8-node GRPO training for DeepSeek V4 Flash on GB200 nodes, with 4 GPUs per node:

The run uses:

```text
model: Deepseek/DeepSeek-V4-Flash-Base
policy backend: AutoModel DTensor/FSDP2
generation backend: vLLM, FP8, DeepGEMM enabled
nodes: 8
gpus_per_node: 4
```

Relevant YAML config

```yaml
grpo:
num_prompts_per_step: 32
num_generations_per_prompt: 16
max_num_steps: 10000
max_response_length: 2048

policy:
model_name: Deepseek/DeepSeek-V4-Flash-Base
train_micro_batch_size: 1
train_global_batch_size: 512
logprob_batch_size: 1
precision: bfloat16
max_total_sequence_length: 4096
logprob_chunk_size: 1024
offload_optimizer_for_logprob: true
dtensor_cfg:
expert_parallel_size: 32
activation_checkpointing: true
moe_parallelizer:
ignore_router_for_ac: true
env_vars:
automodel_kwargs:
backend:
attn: sdpa
linear: torch
rms_norm: torch_fp32
dispatcher: torch
experts: torch_mm
enable_hf_state_dict_adapter: true
generation:
max_new_tokens: 2048
vllm_cfg:
tensor_parallel_size: 8
pipeline_parallel_size: 1
expert_parallel_size: 16
enable_expert_parallel: true
precision: fp8
kv_cache_dtype: fp8_ds_mla
pow2_weight_scaling_factors: true
gpu_memory_utilization: 0.6
max_model_len: 4096
use_deep_gemm: true
colocated:
enabled: true

cluster:
gpus_per_node: 4
num_nodes: 8
```

**Additional context**

This was reproduced multiple times:

```text
2635901: reached step 3, then stopped making progress
2636224: reached step 3, then stopped making progress
2637238: reached step 3, then stopped making progress
2637499: reached step 4, then stopped making progress until cancelled
2638324: debug run showed very slow policy offload/refit and then NCCL timeout
```

The likely hang point is:

```text
DTensorPolicyWorkerV2.offload_after_refit()
-> self.move_to_cpu(self.model)
-> model.to("cpu")
```