patch various small PPO implementation bugs

Author: Jemoka

src/astra_rl/algorithms/ppo.py

@@ -89,6 +89,7 @@ def step(
             torch.tensor(batch.reward)
             .to(logprobs_attacker.device)
             .unsqueeze(-1)
+            .unsqueeze(-1)
             .repeat(1, *values.shape[1:])
         )
         A = Q - values