Change sampling method from randint to choice in Replay and robustify…

… policy networks in SAC (#111)

* Introduced possibility to change alpha parameter

* Fix sum operation which causes trouble for more that two batch dimensions

* Replace randint with choice to avoid duplicates

* Added replace as an option to the replay buffer and a warning if desired behaviour is not possible

Initial v0.2.0 code

Saved version before v0.2 with RIG and HER

Initial version built off of pytorch v0.3

upgrade to MuJoCo 1.5

First tagged version