[recipe] feat: Revamp single-controller demo with agentic multi-turn rollout and add CI

[vermouth1992]

Summary

• Rewrite single_controller_demo.py to showcase a realistic agentic RLHF training loop: replaces the
  previous hardcoded tensor inputs and AsyncvLLMServer with a multi-turn AgentLoop that interleaves
  LLM generation with simulated tool calls, an OpenAI-style MessageDataset with interleaved image
  support, a proper DataLoader-based training loop, and an explicit compute_reward step that writes
  advantages back through TQ.
• Adopt the new KV APIs introduced in [feat,CI] Improve KV API usability and KVBatchMeta interactions #57 throughout the recipe — uses kv_batch_get_by_meta,
  the return value of kv_batch_put (cumulative KVBatchMeta), and removes all manual
  kv_meta.fields.append(...) calls.
• Add structured configuration via @dataclass classes (TrainerConfig, AgentLoopConfig,
  MessageDatasetConfig) and argparse CLI, replacing the flat OmegaConf.create dict. Trainer config
  and TQ config are now cleanly separated.
• Add recipe-check.yml CI workflow that runs the demo end-to-end on every push/PR with reduced
  parameters (--num-samples 8 --global-batch-size 4 --rollout-agent-num-workers 1) to keep CI fast.

Key changes

Area	Before	After
Rollout	AsyncvLLMServer (Ray actor, single-turn)	AgentLoop (multi-turn with tool calls, per-sample async)
Data	Hardcoded [[1,2],[3,4],...] tensors	MessageDataset + DataLoader with random multi-modal messages
Reward	Simulated inline (time.sleep)	compute_reward() producing per-token advantages via TQ
KV API	kv_batch_get(keys=..., fields=...) + manual field tracking	kv_batch_get_by_meta(meta=...) + kv_batch_put return value
Config	Flat OmegaConf dict	Typed @dataclass hierarchy + argparse CLI
CI	None	recipe-check.yml workflow

Test plan

• Run the recipe locally: python recipe/simple_use_case/single_controller_demo.py --num-samples 8 --global-batch-size 4
• Verify the new recipe-check.yml workflow passes in CI
• Confirm existing tests (pytest tests) still pass

[ascend-robot]

CLA Signature Pass

vermouth1992, thanks for your pull request. All authors of the commits have signed the CLA. 👍