KLDiv instead of MSE

[YanickZengaffinen]

In the GT-MBRL paper they use KLDiv as the objective, but we are currently using MSE.

• Implement KLDiv bc it allows loss 0 with random transitions (MSE does not => model would be incentivized to choose short trajectories, as rewards are always negative)

[YanickZengaffinen]

This is what is learned on MDP2 with kldiv. Model reward is 0.0 => it's perfect at predicting the transitions that are taken. Avg Policy Reward on Learned Model: 0.73 => it's nowhere close to the 40 but that doesn't mean much.

More experiments needed (especially: actual reward of policy in real env)