[recipe] feat: Revamp single-controller demo with agentic multi-turn rollout and add CI (Ascend#63)

## 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 Ascend#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

---------

Signed-off-by: Chi Zhang <czhangseu@gmail.com>

Author: vermouth1992

.github/workflows/recipe-check.yml

@@ -0,0 +1,36 @@
+name: Recipe check
+
+on:
+  push:
+    branches:
+      - main
+      - v0.*
+  pull_request:
+    branches:
+      - main
+      - v0.*
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11"]
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v3
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
+        pip install -e ".[yuanrong]"
+    - name: Run recipes
+      run: |
+        export RAY_DEDUP_LOGS=0
+        python3 recipe/simple_use_case/single_controller_demo.py --num-samples 8 --global-batch-size 4 --rollout-agent-num-workers 1