More explicit support for reinforcement learning

returnn-common should come with simple ways to use common building blocks / mechanisms of reinforcement learning, which can be useful in general.

This issue is supposed to be a collection of things we need. Although we probably should have individual issues for each individual feature.

(I consider this as resolved when we have some simple RL examples, which seems simple enough. E.g. producing some common actor critic training example on some common task. Basically reproducing some of the basic examples of other RL frameworks. Only when we have that, we know that returnn-common covers the basic needed utilities.)

rwth-i6 / returnn_common

More explicit support for reinforcement learning #19

20