Make run_experiment more modular

[mctigger]

https://github.com/cpnota/autonomous-learning-library/blob/3ec67d5c05c521344bbba47045c1b179a8ccc275/all/experiments/run_experiment.py#L5-L13

run_experiment should only take agents, envs and experiment. Then we could call it like this:

def make_experiment(agent, env):
    return CustomExperiment(
        agent, env, render=False, quiet=False, write_loss=True # Allows for any interface
    )

run_experiment(
         agents, envs, make_experiment
)

[michalgregor]

I think it would also be useful if run_experiment returned something - possibly the experiment object itself. Useful e.g. when you want to visualise the curves from the tensorboard log. Right now you don't even know what directory they're in if you are using run_experiment, right?

[cpnota]

Will address as part of #290