On a few pitfalls in KL divergence gradient estimation for RL