Add LSTM policy

[kim-mskw]

To enhance the RL method currently implemented in ASSUME, we want to add diffrent policies and buffers.

This includes an LSTM policy for the actor and/or critic instead of the plain NN to have some time awareness.

Furthermore, a new buffer is implemented, which uses not random sampling but importance sampling instead.

[nick-harder]

As mentioned by @kim-mskw during the development meeting, the work on the LSTM policy provided good results and can be merged with the framework soon