Feature/save agents

[cpnota]

It's here! This PR dds the ability to save/load agents, addressing #161 .

There are a few keys to the design. First of all, rather than having agent.act(state) and agent.eval(state), agents were split into a training-mode Agent and a TestAgent. Both types of agents can be instantiated by a Preset:

agent = preset.agent()
test_agent = preset.test_agent()

The Preset is a serializable object containing the hyperparameters and all necessary torch models. The TestAgent inherits a copy of the model trained by the Agent, allowing TestAgents from different points in training to be stored.

The second major key to the design is that the Preset, rather than the Agent itself is saved:

preset.save(filename)
preset = torch.load(filename)

This is important because the underlying Agent objects are often difficult to serialize, and even if they can be serialized they can take up an excessive amount of storage (for example, a standard 1 million frame Atari replay buffer is ~7 GB).

One thing to note is that while this design supports creating a training mode Agent with a previously trained network, it does not support a full "resume" of training, e.g., scheduler states will be reset and replay buffers will be cleared. Full resume functionality introduces many difficulties which interfere with the design of the library, however, we may implement a partial solution in the future.

Example usage can be found below:

# construct the preset
preset = builder().hyperparameters(lr=1e-3).env(some_env).build()

# run agent in train mode
agent = preset.agent()
for i in range(episodes)
    run_episode(agent, some_env)

# run agent in test mode
test_agent = preset.test_agent()
for i in range(test_episodes)
    run_episode(test_agent, some_env)

# save the model for later
preset.save('dqn.pt')

# load model from disk and watch
preset = torch.load('dqn.pt')
test_agent = preset.test_agent()
for i in range(test_episodes)
    run_episode(test_agent, some_env, render=True)