MuJoCo environments

[tristandeleu]

• MuJoCo environments (HalfCheetah and Ant) now inherit from gym
• Add normalization wrappers: NormalizedActionWrapper, NormalizedObservationWrapper and NormalizedRewardWrapper
• Bump the versions of the MuJoCo environments to -v1. The new versions of these environments have normalized actions
• Fix normalization of the advantages in the outer loss
• Weights initialization: the weights of the models are now initialized with xavier_uniform
• Variance of NormalMLPPolicy is now a parameter, independent of the input