[infra] Skip extra forward pass when policy loss does not require `old_action_log_probs`

[erictang000]

This happens for async RL or in general when mini_batch_size == train_batch_size - we unnecessarily do a fwd pass on the old policy model even though the policy loss may not use it.