[Performance] Avoid cloning trajs in SliceSampler

ghstack-source-id: 2e133fcea716b202694cfa84df3f6e4ba3507bbc
Pull Request resolved: #2671

torchrl/data/replay_buffers/samplers.py

@@ -1243,7 +1243,7 @@ def _get_stop_and_length(self, storage, fallback=True):
                             "Could not get a tensordict out of the storage, which is required for SliceSampler to compute the trajectories."
                         )
                 vals = self._find_start_stop_traj(
-                    trajectory=trajectory.clone(),
+                    trajectory=trajectory,
                     at_capacity=storage._is_full,
                     cursor=getattr(storage, "_last_cursor", None),
                 )