feat(dynamo): add native token wrapper transport

[jthomson04]

Summary

• add a NeMo-RL HTTP token wrapper in front of Dynamo chat completions when the Dynamo generation path exposes an OpenAI-compatible rollout server
• have the wrapper render chat prompts to token IDs, send them to Dynamo via nvext.token_data, and request only nvext.engine_data
• require successful Dynamo chat responses to include native nvext.engine_data.prompt_token_ids and completion_token_ids, so Gym does not fall back to Dynamo /tokenize
• share prefix-substitution logic between the existing vLLM worker path and the Dynamo wrapper
• pass tokenizer state into Dynamo generation from GRPO, while keeping direct /completions calls pointed at the real Dynamo frontend
• update DGD worker manifests to launch Dynamo workers with --enable-rl and handle newer rl worker discovery/pause semantics for MX refits
• update the Gym submodule pointer for native Dynamo engine_data token consumption; companion Gym PR: feat(vllm-model): consume native dynamo token data NVIDIA-NeMo/Gym#1784

Validation

• git diff --cached --check
• uv run ruff check nemo_rl/algorithms/grpo.py nemo_rl/models/generation/dynamo/dynamo_generation.py nemo_rl/models/generation/dynamo/token_wrapper.py nemo_rl/models/generation/vllm/vllm_worker_async.py nemo_rl/utils/prefix_reuse.py tests/unit/models/generation/test_dynamo_generation.py tests/unit/models/generation/test_dynamo_token_wrapper.py tests/unit/models/generation/test_vllm_generation.py tests/unit/utils/test_prefix_reuse.py
• uv run pytest tests/unit/models/generation/test_dynamo_token_wrapper.py tests/unit/utils/test_prefix_reuse.py tests/unit/models/generation/test_dynamo_generation.py -q -> 44 passed, 1 warning
• uv run pytest tests/unit/models/generation/test_vllm_generation.py -q -k replace_prefix_tokens -> 5 passed, 37 deselected
• in 3rdparty/Gym-workspace/Gym: uv run pytest responses_api_models/vllm_model/tests/test_app.py -q -k TokenIDInformation -> 3 passed, 66 deselected
• in 3rdparty/Gym-workspace/Gym: uv run ruff check responses_api_models/vllm_model/app.py responses_api_models/vllm_model/tests/test_app.py

[github-actions]

✅ Submodule Fast-Forward Check Results

Check based on commit: 94d4852 (PR #9 from codex/dynamo-native-token-wrapper)

✅ Submodules that are properly updated:

Gym: ✅ PR branch is ahead of dynamo-k8s-integration branch (fast-forward)

All submodule changes look good! ✨