open_mfdataarray with float-type tolerance raises exception

[martijnende]

Calling open_mfdataarray (or similar) with e.g. tolerance=1.0 raises:

File ~/miniconda3/envs/nami/lib/python3.12/site-packages/xdas/core/routines.py:731, in concatenate(objs, dim, tolerance, virtual, verbose)
    729 if dim_has_coords:
    730     if hasattr(coord, "simplify"):
--> 731         coord = coord.simplify(tolerance)
    732     else:
    733         if tolerance is not None:

File ~/miniconda3/envs/nami/lib/python3.12/site-packages/xdas/core/coordinates.py:854, in InterpCoordinate.simplify(self, tolerance)
    852     else:
    853         tolerance = 0.0
--> 854 tie_indices, tie_values = douglas_peucker(
    855     self.tie_indices, self.tie_values, tolerance
    856 )
    857 return self.__class__(
    858     dict(tie_indices=tie_indices, tie_values=tie_values), self.dim
    859 )

File ~/miniconda3/envs/nami/lib/python3.12/site-packages/xdas/core/coordinates.py:1049, in douglas_peucker(x, y, epsilon)
   1047 dmax = d[index]
   1048 index += start
-> 1049 if dmax > epsilon:
   1050     stack.append([start, index + 1])
   1051     stack.append([index, stop])

UFuncTypeError: ufunc 'greater' did not contain a loop with signature matching types (<class 'numpy.dtypes.TimeDelta64DType'>, <class 'numpy.dtypes._PyFloatDType'>) -> None

I guess that douglas_peucker really wants epsilon to be a timedelta. A quick fix would be to promote epsilon to a np.timedelta64 object here, or otherwise manipulate tolerance here.

Note, this behaviour could be specific to the NumPy version (== 2.3.5), which may have permitted timedelta-float comparisons in previous versions.