use hyperparameters from vectorized run to test training of vectorized and non-vectorized
    vectorized hyperparams: check
    vectorized training seems to work

checklist
    non-vectorized hyperparameters for ppo DiscreteSteps-v0 - DONE
        do non-vectorized hyperparameters work for vectorized?
            if yes, add commented vec_env_wrapper to tuned DiscreteSteps-v0 params
        are non-vectorized hyperparameters similar to vectorized ones? - NOPE
    vectorized lstm hyperparameters for unstacked DiscreteSteps-v0 - DONE
    non-vectorized hyperparameters for ppo ContinuousJumps-v0

make gif of single best agent, not both discrete and continuous
    --last agent in ppo discrete steps 5

later

environment factories, duplication in custom_environments is a big nono

render:
Would other memory types be faster?
Would cuda/numba fastmath be faster?
Use numba.jit for GPU / CPU compatibility
