Fix PrioritizedReplayBuffer filtering

[alexandrelarouche]

Fix error where add() crashes if no trajectory is better than buffer trajectories.

Specifically, the following lines gave results which couldn't be handled because the Trajectories object was empty:

                batch = training_objects.last_states.tensor.float()
                batch_dim = training_objects.last_states.batch_shape[0]
                try:
                    batch_batch_dist = torch.cdist(
                        batch.view(batch_dim, -1).unsqueeze(0),
                        batch.view(batch_dim, -1).unsqueeze(0),
                        p=self.p_norm_distance,
                    ).squeeze(0)

Fix is trivial: flow control out of the function if no trajectories should be added to the buffer.

[josephdviviano]

awesome - thanks for the fix!