Deep Reinforcement Learning Methods - Week 8

This week I implemented async n-step q learning and async one step sarsa. Both fit into my async learning framework very well. Although those two async algorithms, as well as async q learning and upcoming a3c, share many common code snippets, I find it very difficult to unify them further. I tried that in python, it ended up with many if clauses, which heavily hurt the readability of the code. And it will be even more troublesome if we want to support LSTM layers. In the meantime I also worked on the PR of async one step q learning. We introduced HAS_OPENMP to be compatible with non-openmp compiler. We also totally eliminated the explicit use of lock by a trick of reordering push and pop and thus we only need omp critical, which I think made the code more clear. I think this PR now is finally ready to merge.