fix(vlm): forward image payloads in async rollouts

[jinglinglingling]

What does this PR do ?

Preserves native multimodal side-channel fields when SingleController dispatches async VLM rollouts, so vLLM and the training forward receive the same image input.

The async per-sample path previously rebuilt generation input from tokenized message data but dropped fields such as vllm_content and vllm_images. This made vLLM generation image-blind while Megatron recomputed log probabilities with the image, producing large token_mult_prob_error values.

For multi-turn rollouts, later turns retain native media while clearing stale vllm_content, allowing the updated tokenized prefix to be used.

This PR does not modify recipes or generation lengths. The existing defaults remain Qwen max_new_tokens=1024 and Nemotron max_new_tokens=4096.

Validation results:

• Controlled unfixed Qwen reproduction: train/token_mult_prob_error = 2.77639
• Same controlled setup after the fix, vLLM TP1/TP2/TP4: approximately 1.02
• Native Qwen configuration (max_new_tokens=1024, 8 prompts x 16 generations), 100/100 steps: all TMPE values below 1.05; range 1.02006–1.02614; truncation rate 0
• Native Qwen W&B run: https://wandb.ai/nvidia/rohit-async_grpo_v2/runs/4t6osucc
• Native Nemotron configuration (max_new_tokens=4096, 8 prompts x 16 generations), 100/100 steps completed. Most aggregate TMPE values were approximately 1.03; three outliers remain under investigation, including one immediately after checkpoint resume.
• Native Nemotron W&B runs: https://wandb.ai/nvidia/rohit-async_grpo_v2/runs/5g2qyqmb and https://wandb.ai/nvidia/rohit-async_grpo_v2/runs/y2govxqy

Reproduction commands

Qwen native-length 100-step validation:

uv run --locked examples/run_grpo_single_controller.py \
  --config examples/configs/recipes/vlm/vlm_grpo-qwen2.5-vl-3b-instruct-clevr-1n8g-megatron-single-controller-async.v1.yaml \
  grpo.max_num_steps=100 \
  logger.wandb_enabled=true \
  logger.tensorboard_enabled=true \
  logger.monitor_gpus=false \
  logger.wandb.project=rohit-async_grpo_v2 \
  logger.wandb.name=linglinj-qwen-image-fix-native1024

Nemotron native-length 100-step validation (3 nodes / 8 GPUs per node):

uv run --locked examples/run_grpo_single_controller.py \
  --config examples/configs/recipes/vlm/vlm_grpo-nemotron-omni-30ba3b-clevr-3n8g-megatron-tp8ep8-single-controller-async.v1.yaml \
  grpo.max_num_steps=100 \
  policy.train_global_batch_size=128 \
  logger.wandb_enabled=true \
  logger.tensorboard_enabled=true \
  logger.monitor_gpus=false \
  logger.wandb.project=rohit-async_grpo_v2 \
  logger.wandb.name=linglinj-nemotron-native4096-gbs128

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[jinglinglingling]

/ok to test 2bb044e

[rohitrango]

/ok to test 167ad7b

[rohitrango]

/ok to test 188c841

[jinglinglingling]

Addressed the remaining review comment with a two-turn _run_single_rollout regression test in ec12bc64. The PR is labeled CI:L2; @rohitrango, could you approve the latest SHA with /ok to test ec12bc64 so the L2 suite can run again?

[rohitrango]

could you sign your latest commit? once it passes DCO I will re-run

[rohitrango]

/ok to test 2037327

[rohitrango]

/ok to test 6d89cac