v1.2 Actor Critic

• use actor-critic as RL framework
• fix reward that is always negative resulting in wrong convergence with log probabilities

Benchmarks will come soon, don't use trained models of this release

v1.1

• benchmarks on MNIST & RESISC-45