Remove GaussianPolicy and obsolete policies

[muupan]

Resolves #276 and #281

Since LinearGaussianPolicyWithDiagonalCovariance hard-codes squashing actions by tanh, replacing it with FCGaussianPolicy will change the behavior. I confirmed that train_a3c_gym.py can solve InvertedPendulum-v1 as before by running python train_a3c_gym.py 4 --env InvertedPendulum-v1 --arch LSTMGaussian --t-max 50.

[prabhatnagarajan]

LGTM;