Polars seems to allow it, except null is preserved, is that what you mean to ban?

from datetime import datetime
import polars as pl

pl.Series([
    datetime(year=2020, month=1, day=1, hour=0, second=0, minute=0),
    None
]).cast(pl.Int64)

shape: (2,)
Series: '' [i64]
[
        1577836800000000
        null
]

yeah I think so, I think it'd be better to nudge people towards .dt.timestamp / .dt.microsecond / etc. depending on which they want, it should be more robust than casting

Do we still want this?

yes I think we should raise where possible when trying to cast a datetime to a numeric type. At least for pandas-like and PyArrow (those are the ones people are most likely to test, so if they find that it raises in those cases, they'll be less likely to use it)