Improved datetime format inference

[LeoGrin]

Fix #540
The bug in #540 was related to pandas 2.0 update: as datetime format inference became stricter, a bug which was silent (inconsistent date format inferred for one column in the test) became loud.

This PR fixes this and improves date format detection (and thus datetime column detection in TableVectorizer). This replaces using pd.datetime directly for datetime format inference.

To do this, we try pandas's guess_datetime_format (on a subset) both with dayfirst being True and False, and see if one of these options finds a single format for all rows. If both work, we return the monthfirst format and raise a warning. If both return multiple formats, we give up.

This finds the right %d-%m-%Y format for the failing test example, instead of failing or returning mixed format (see below).

  "11-12-2029",
  "02-12-2012",
  "11-09-2012",
  "13-02-2000",
  "10-11-2001"

pd.to_datetime format inference

If I understand correctly, pandas datetime format inference (in pd.to_datetime) works like this:

Versions < 2.0

If not infer_datetime_format, the format of each row is inferred independently.
If infer_datetime_format, the format is inferred from the first non-missing example, and pandas tries to apply it to the other example (but can use another format otherwise).
This can easily create issues, for instance in our tests:

    "11-12-2029",
    "02-12-2012",
    "11-09-2012",
    "13-02-2000",
    "10-11-2001"

For all rows, the inferred format was %m-%d-%Y expect 13-02-2000, for which the format was %d-%m-%Y.

Version >= 2.0

The format is inferred from the first non-missing example, and pandas tries to apply it to the other example (and raise an error otherwise).

[LilianBoulard]

Thanks for the PR, a few tweaks to merge and it's good for me!

[LilianBoulard]

Maybe a remark on the _infer_date_format function: since it's only used in the TableVectorizer, I think it should be moved over to its file.

[LilianBoulard]

Messed up something in my previous review, this is a fix

[LilianBoulard]

LGTM, thanks again!