Constraint-Procedural Logic Generated Environments for Deep Q-learning Agent training and benchmarking