Avoid generic-unit timedelta in CDFepoch.to_datetime (fixes #333)

[gaoflow]

Summary

CDFepoch.to_datetime raises a DeprecationWarning under NumPy 2.5+:

DeprecationWarning: The 'generic' unit for NumPy timedelta is deprecated, and will
raise an error in the future. This includes implicit conversion of bare integers
(e.g. `+ 1`). Please use a specific unit instead.

Reproducer from #333:

import warnings
warnings.simplefilter('always')
from cdflib.epochs import CDFepoch
CDFepoch.to_datetime(631377279184000000)   # warns twice

Cause

In _compose_date (cdflib/epochs.py):

• total_datetime = np.array(sum(arrays)) — sum starts from the integer 0, so the first operation is 0 + datetime64[...]. Adding a bare integer to a datetime64 yields a generic-unit timedelta, which NumPy now deprecates.
• np.where(nat_positions, np.datetime64('NaT'), total_datetime) — the bare np.datetime64('NaT') is also generic-unit and triggers the same warning when combined with the typed total_datetime.

Fix

• Sum the components starting from the first array (a datetime64) instead of the implicit 0.
• Build NaT with the same unit as the data (np.datetime_data(total_datetime.dtype)[0]).

Both are behaviour-preserving — only the intermediate dtypes change, not the result.

Verification

• Reproduced the two warnings on a NumPy nightly (2.6.0.dev); after the fix, to_datetime emits no generic-unit DeprecationWarning and returns the identical value.
• Full test suite green on released NumPy 2.4.6 (82 passed); the epoch tests pass on both released and nightly NumPy.
• Added test_to_datetime_no_generic_timedelta_warning, which fails on the current code under the nightly and passes with the fix (and passes trivially on older NumPy that doesn't emit the warning).