Adapt to deprecation of generic np.timedelta64 dtype

[spencerkclark]

This PR modifies the few places we relied on a generic np.timedelta64 dtype to explicitly specify the time resolution:

• It removes NAT_TYPES and relies instead on checking the dtype.kind in computation.nanops._maybe_null_out.
• It infers the time unit using np.datetime_data from the input dtype to determine the unit on the returned fill_value in core.dtypes.maybe_promote.
• It explicitly constructs a zero-valued np.timedelta64 or np.datetime64 object for use downstream in plot.utils._determine_cmap_params.

[keewis]

looks good to me. If you're up for it, the same warning is also raised in test_concat.py::test_concat_type_of_missing_fill and test_dtypes.py::test_maybe_promote[m-expected14].

[spencerkclark]

Oops thanks @keewis for noting those additional related failures I missed. This should be good for another look.