fix(profiling): use SGD for pi0/pi05/pi0_fast and free CUDA cache after deterministic forward

Adam optimizer states (exp_avg + exp_avg_sq) require ~16GB extra on top of
model params and gradients for 4B parameter models, exceeding the 22GB GPU.
SGD has zero optimizer state overhead and profiling only measures
forward/backward timing anyway.

Also adds torch.cuda.empty_cache() after deterministic forward to release
transient memory before the training loop starts.

Made-with: Cursor

src/lerobot/utils/profiling_utils.py

@@ -343,6 +343,8 @@ class TrainingProfiler:
             output_dir=self._output_dir,
             device_type=self._device.type,
         )
+        if self._device.type == "cuda":
+            torch.cuda.empty_cache()
 
     def __enter__(self) -> TrainingProfiler:
         if self._device.type == "cuda":