Add warning when user tries to train a model using datetime values without a datetime_format set

[srinify]

Situation

If you read data into a pandas dataframe, datetime columns often are (by default) set to the object (aka string) dtype. The user then must take some action to make SDV synthesize better datetime values:

• user can manually cast the column to datetime dtype
• user can set a datetime_format when creating the metadata for SDV to use

It's easy for someone (especially a new user) to skip this step entirely, causing issues later.

Suggested Warning

After discussing with Neha, I'm opening this feature request. Ideally, we can add a warning when the user tries to make progress with SDV (e.g. maybe training a synthesizer) that they should add a datetime_format.

[npatki]

Note that metadata auto-detection will generally pick up a datetime_format for most common cases. But there are other ways of creating metadata, for example manually writing a Python dict or JSON file. In such cases, there may not be a datetime_format.

This warning should only appear when:

• In metadata, thesdtype is 'datetime' AND
• In metadata, there is no datetime_format specified AND
• In the data, the dtype (storage type) is 'object'

Suggested API

This warning should result from metadata.validate_data function. Ideally the warning can pretty-print a list of columns to watch out for.

>>> metadata.validate_data(data)
Warning: No 'datetime_format' is present in the metadata for the following columns:

Table Name    Column Name    sdtype    datetime_format
users         start_date     datetime  None
users         end_date       datetime  None
sessions      timestamp      datetime  None

Without this specification, SDV may not be able to accurately parse the data. We recommend adding datetime formats using 'update_column'.

Note: For single table, we can exclude the Table Name.