refactor: emit strategy metrics as floats

- Converts detached strategy metrics before returning loss output
- Removes redundant item conversion from the trainer loop
- Updates the documented contract and regression tests

docs/guides/training.md

@@ -89,7 +89,7 @@ on_train_end
 
 Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric` (JSONL + validation, rank-0), `progress_bar` (tqdm), `gradient_clipping` (always registered; computes grad norm, clips only when `max_grad_norm` is not `None`).
 
-Strategies return `{"loss": Tensor, "metrics": Dict[str, Tensor]}` when called by the trainer. Built-in metrics include the task-specific loss and, for MoE models, `moe_aux_loss` plus `moe_aux_loss_weighted`. Direct `compute_loss(batch)` calls continue to return a single loss tensor.
+Strategies return `{"loss": Tensor, "metrics": Dict[str, float]}` when called by the trainer. Built-in metrics include the task-specific loss and, for MoE models, `moe_aux_loss` plus `moe_aux_loss_weighted`. Direct `compute_loss(batch)` calls continue to return a single loss tensor.
 
 ## Strategies