refactor: unify rollout onto inference engine KV-cache path

- RolloutGenerator now delegates prefill/decode to InferenceScheduler.run_batch (sync API, no background thread), sharing one KV-cache code path with the inference server and eliminating O(n^2) recompute in rollout
- Add sample(return_logprobs=) and Executor.execute_decode(return_logprobs=) to expose behaviour-policy log-probs through the engine; Task gains output_logprobs
- RolloutResult now subclasses RawRollout (adds rewards only), removing duplicated fields
- RolloutRunner.__call__ returns (result, is_fresh) instead of relying on object identity, removing the fragile refresh-detection contract
- Remove O(n^2) generate_responses helper and dead code (_tokenize_prompts, unused old_model arg)
- train_context.py wires InferenceScheduler directly instead of hand-rolling SamplingPipeline
- Tests: +11 covering return_logprobs, run_batch, and KV-cache-backed rollout semantics; 404 pass

tests/trainer/test_online_strategy.py

@@ -71,22 +71,31 @@ def _make_rollout_result(B=2, G=4, P=6, R=8, device="cpu"):
 
 
 class _RecordingRunner:
-    """Fake RolloutRunner that returns a fixed result and tracks calls."""
+    """Fake RolloutRunner returning a fixed result with freshness tracking.
+
+    Freshness is ``True`` on the first call after construction or after
+    :meth:`swap_result`; ``False`` on subsequent cached calls — mirroring
+    the real ``RolloutRunner`` contract without invoking generation.
+    """
 
     def __init__(self, result):
         self.result = result
         self.calls = 0
         self.step_calls = 0
+        self._fresh = True
 
     def __call__(self, batch):
         self.calls += 1
-        return self.result
+        fresh = self._fresh
+        self._fresh = False
+        return self.result, fresh
 
     def step(self):
         self.step_calls += 1
 
     def swap_result(self, result):
         self.result = result
+        self._fresh = True
 
 
 @pytest.fixture