Should we only do this for pd.date_range for now, and not yet for pd.timedelta_range, since in general timedelta still defaults to nanoseconds? (eg when converting strings or integers in pd.to_timedelta)

I think best case would be to do it all at once (xref #63018 gets everything but the string-parsing inference for td64). But if it'll make a difference in getting a release out soon, I'll happily revert.

Ah, I hadn't seen #63018. Yes, my preference would also be to do it all at once. Will try to take a look at that one then.