Feature dataset checkpoint strategy

[plaguss]

Description

This PR modifies enable_checkpoint in the Pipeline.generate to allow a more behaviour during the generation (saving while generating).

The following example allows saving the CustomDataset every 100 generations to disk.
By default, it will use DatasetCheckpoint(save_frequency=-1), which replicates the previous behaviour of enable_checkpoints=True.

from distilabel.utils.dataset import DatasetCheckpoint

pipe.generate(
    ...,
    dataset_strategy=DatasetCheckpoint(save_frequency=100)
)

Closes #168 and #193.

[gabrielmbmb]

Hi @plaguss! Much needed feature :) it looks good, I left some comments

[plaguss]

Added a small reference in the docs, inside a new section of the Pipeline from the concept guides

[alvarobartt]

LGTM even though I'm still pending to test it myself!

Also a general comment is that I'd like to discuss on the naming with the rest of the team, as IMO we could just have checkpoints_path and checkpoints_freq or something similar, rather than asking the user to instantiate a class, while we may use/re-use the class internally, but to avoid the user for instantiating multiple classes within the same function.

[ignacioct]

Just a couple of personal doubts and minor suggestions, but looks really nice.