fix: support string, date, and datetime bounds in polars fill_gaps

[Jakub Krasuski (Blizzeq)]

When calling fill_gaps on a Polars DataFrame with a custom start or end timestamp (such as a string like '2020-01-01', a date, or a datetime), _determine_bound_pl created a Series([val]) without matching the target dtype of the time column. This caused a panic in datetime_ranges when passed string series, or a schema error when joining differing datetime and date keys.

This PR ensures _determine_bound_pl converts scalar bounds (strings, dates, datetimes) to match the time column's target dtype in Polars before building the grid. Added regression tests covering string, date, and datetime bounds on Polars.

[José Morales (jmoralez)]

Hey Jakub Krasuski (@Blizzeq), thanks for your interest in this issue. We try to mimic what each backend supports so pandas allows using strings as dates/datetimes but polars always uses date and datetime objects. we used to have this distinction in the docs but they were migrated to the tests.

What I'd prefer here would be to add a check in the polars branch to see if the bound and the time column don't match and raise a clear error there, e.g. bound is datetime but time column is date.

[Jakub Krasuski (Blizzeq)]

Makes sense, thanks. I dropped the conversion and added a check in the polars branch instead: start and end now have to match the time column, otherwise fill_gaps raises a ValueError that names the argument and the expected type.

[José Morales (jmoralez)]

Hey Jakub Krasuski (@Blizzeq), I'm very sorry but on a second thought I prefer your first approach. Can you please revert the latest commit so we can pick up from there?

[Jakub Krasuski (Blizzeq)]

No problem, reverted in 19dda74 so the branch is back to the conversion approach from the first commit.