fix: use sum reduction for DPO sequence logprob

- DPO requires sequence-level sum of token logprobs, not per-token mean
- mean reduction made beta*ratio_diff ~0.03 (near-zero gradient)
- loss stalled at 0.6931 because logsigmoid(0.03) has vanishing grad
- sum gives beta*ratio_diff ~10 with meaningful gradients

astrai/trainer/strategy.py

@@ -225,7 +225,7 @@ class DPOStrategy(BaseStrategy):
         device: str,
         ref_model: nn.Module,
         beta: float = 0.1,
-        reduction: str = "mean",
+        reduction: str = "sum",
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)