[FEAT] Use dates for train/val/test split

[marcopeix]

Add the option to use dates instead of number of time steps to perform the train/val/test split in cross-validation.

[elephaint]

I get the intuitive part 😄 .... but:

• Input validation is tough:
  • What if df doesn't have timestamps
  • What if validation date is later than test date
  • Different timestamp formats
• Kinda breaks API consistency across all the repos (we'd have to add it in cv endpoints sf / mf / nixtla too) - doesn't mean we shouldn't implement it, but just noting.
• val_size and test_size are used quite extensively throughout the library; if we support this in the cv endpoint, would make sense to make it consistent and support it in all endpoints where test_size and val_size are provided as an argument.

If kept very consistent throughout NF and strong input validation, we could use it, but it remains hard....

E.g. "2011-03-02", how does it know what date that is if the format isn't supplied? So, it should then check against the format of the df I guess?

[elephaint]

Allowing Any here is a validation nightmare, especially with dates that can be timestamp, pandas offsets, strings....

[marcopeix]

I get the intuitive part 😄 .... but:

* Input validation is tough:
  
  * What if df doesn't have timestamps
  * What if validation date is later than test date
  * Different timestamp formats

* Kinda breaks API consistency across all the repos (we'd have to add it in cv endpoints sf / mf / nixtla too) - doesn't mean we shouldn't implement it, but just noting.

* val_size and test_size are used quite extensively throughout the library; if we support this in the cv endpoint, would make sense to make it consistent and support it in all endpoints where test_size and val_size are provided as an argument.

If kept very consistent throughout NF and strong input validation, we could use it, but it remains hard....

E.g. "2011-03-02", how does it know what date that is if the format isn't supplied? So, it should then check against the format of the df I guess?

I agree, and I'm not sure if I like this feature either haha 😅. I knew validation would be hard, and I still find it fragile. To be honest, it hasn't been a wildly requested feature. Let's store it aside for now, and see if it's something the community really wants.