Improve grouped-actions performance

[Max-We]

The grouped-actions wrapper performance is lower compared to other environments (e.g. https://github.com/uvipen/Tetris-deep-Q-learning-pytorch) by rougly 5x. The goal would be to equally performant or better than this baseline.

As a first step, one could measure the performance (SPS) for a random policy without learning for this project and the baseline and investigate potential bottlenecks.