[feature request] Low-level API

Having a lower level API can be helpful for some users to have more control over the training loop. The low-level API is particularly useful in Hierarichal RL or Multi-Agent RL problems. One idea is to provide an API like this repo from @eleurent:

for episode in episodes:
    while not done:
        action = agent.act(state)
        next_state, reward, done, info = env.step(action)
        agent.record(state, action, next_state, reward, done, info)

[Miffyli]

Agreed, this question came up in sb repository quite often. Another related thing we could do is to make getting action probabilities/values bit easier / clarified.

[mhtb32]

Thanks. I think it is worth adding a tutorial about using low-level API to docs.