Skip the cold-start SFT best ckpt and free CUDA cache after eval

Step 0 generate was writing a 5GB success=0 snapshot and leaving the
32GB card fragmented, so the next Adam step OOM'd after batch-32 eval.
Only promote _best after step 0 and empty_cache when eval returns.

train_sft.py

@@ -348,7 +348,7 @@ def main() -> None:
                             },
                             step=step,
                         )
-                    if _is_better_eval(
+                    if step > 0 and _is_better_eval(
                         printable["success_rate"],
                         printable["chrf"],
                         best_success,
@@ -362,6 +362,8 @@ def main() -> None:
                             f"chrf {best_chrf:.2f} -> {best_path}"
                         )
                     model.train()
+                    if device == "cuda":
+                        torch.cuda.empty_cache()
             if ckpt_now:
                 _save(last_path, dump())
                 print(f"   last -> {last_path}")