Unify restoring variables

[jindrahelcl]

As I looked through the code, in tf_manager.__init__(), there is an option to supply initial variable files. There is a similar option in [main] section of the config. I did not investigate if the two are colliding in any way, but I find it strange at least.

So, my preferred solution is to throw the initial_variables out of learning_utils and leave only those in tf_manager.

With the refactoring for TF 1.0, I will move all remaining saving and restoring functionality from learning_utils to tf_manager anyway. (Including the decision whether to save the variables according to their score.)

[jlibovicky]

Sounds reasonable.