fix: support datetime variables in Dataset.interp

[emmanuel-ferdman]

PR Summary

Dataset.interp() silently dropped datetime64 and timedelta64 variables. Now they are interpolated by converting to float64 and back, with NaT handled like NaN.

• Closes Dataset.interp() silently dropping time-like data arrays #10900
• Tests added

[spencerkclark]

Thanks @emmanuel-ferdman—this largely looks good.

Could you add "Mm" to the valid dtypes on the DataArray.interp side as well? There is no need now to raise in that circumstance, which is great.

One edge case to consider is what to do in the scenario that extrapolation leads to float values outside the range that can be represented by 64-bit integers. Maybe we punt on that for now though, since it could be messy to handle in a robust way.

[emmanuel-ferdman]

@spencerkclark Thanks, updated :) Regarding the offset-based approach, I think we could subtract the minimum datetime before converting to float, which gives nanosecond precision - but it would break dask laziness since we'd need to compute .min().

[spencerkclark]

Thanks @emmanuel-ferdman! This looks good from my perspective.