Add the capability to sample trajectories

[jamartinh]

For some algorithms it is very useful to be able to sample random trajectories of max_len to take advantage of the temporal series information.

For instance, rb.sample_trajectories(n_samples: int, trajectory_max_len: int)

So you get a collection of samples each one containing an ordered part of an experienced episode

[ymd-h]

@jamartinh

Thank you for your feedback.

Since we couldn't define agreeable standard usage, we don't provide a specific class for episodic experiment (at least now),
however, I still believe we can build an episodic configuration with current (Prioritized)ReplayBuffer
by setting a trajectory (aka. episode) as a single record (= step in non-episodic experiment).

If the trajectory is always same length (= T), the situation becomes simple.
At the env_dict parameter of the constructor, you can use (T, *original_shape) or (*original_shape, T) for the episodic values.
Then you can add whole trajectory at once.
If you need temporary storage for ongoing current episode, you might use another ReplayBuffer.

If the trajectory length differ each time, I think there are 2 options.

1st option: Use padding
As far as I know, anyway we pad data before passing neural network, so that it makes sense for me to pad trajectory before storing.
This might requires unnecessary memory.

2nd option: Use dtype=object
By specifying dtype as object (with shape as 1) , Python object are stored.
Python object is a kind of pointer, so that it is also possible someone might change the actual data.

I hope it will help you.