Add a loop to run SGD for several epochs for each update, which can make the training process faster

[yonatank93]

In the current version, the agent only perform a single gradient update for every train_frequency, which means that gradient update is not really done frequently. This may results in slow training.

If we allow multiple gradient updates for every train_frequency, the agent may be able to learn more quickly. So, we should add this option. Especially, since SB3 also has this option.

However, we also cannot do too many gradient updates, as the experience buffer, where the training data are sampled from, keeps changing. Too many gradient updates would cause the agent to overfit to specific buffer versions.

[yonatank93]

Also, allow specifying model architecture, which is currently hard-coded into agent.py.

[yonatank93]

And add an option to specify the model file when instantiating the agent.
I think I previously suggested to remove this option, but now I see that if we include this option, user can load the model from the configuration file directly in one line, instead of needing to instantiate the agent then loading the model in 2 separate steps.

[yonatank93]

We should resample the buffer at every gradient step.