FEAT: Training window filtering

[marcopeix]

Currently, we create the maximum number of training windows, meaning that we might have windows with only 1 available insample data point and 1 available outsample data point.

These are technically low quality windows.

This PR adds the parameter available_sample_fractions to control how many available insample and outsample data points should be available as a fraction of input size and horizon for insample and outsample respectively.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[marcopeix]

It's basically a rework of #1059 . Many thanks to @jasminerienecker for the initial idea!

[elephaint]

Nice work, few comments!

[elephaint]

Thanks, good work! 2 things:

• It needs a proper test (the current one has no value)
• Minor adjustment in the implementation is required for multivariate models I think (which becomes apparent in a test)

[elephaint]

Great work! Good for me if you agree with the final changes I made (cosmetic and I changed the tests + added explanation to the tests)

[elephaint]

@marcopeix good to go once you verify/check my explanations in the test

Incorrect

[Antoine-Schwartz]

Sampling quality is a really interesting subject. Thank you for taking the time to add this first option @marcopeix !

By the way, it would be great to have a dedicated documentation section.
Sampling with time series, and particularly the neuralforecast implementation with the mask system is not obvious to understand.
And if you also look at the interactions with other options such as start_padding_enable, you can quickly get lost :)