Fix dask statistics

[keflavich]

Dask statistics were returning nans if any chunk was all-nan. This fixes that issue, at a slight computational cost (I hope slight, anyway).

Example that I put in a regression test (approximately):

>>> cube = SpectralCube.read('tests/data/example_cube.fits', use_dask=True)
>>> cube._data[:,:,:2] = np.nan
>>> cube.rechunk((1,2,2)).statistics()
{'npts': 28.0,
 'min': <Quantity nan Jy / beam>,
 'max': <Quantity nan Jy / beam>,
 'sum': <Quantity 0.09592362 Jy / beam>,
 'sumsq': <Quantity 0.00211593 Jy2 / beam2>,
 'mean': <Quantity 0.00342584 Jy / beam>,
 'sigma': <Quantity 0.00813614 Jy / beam>,
 'rms': <Quantity 0.00869304 Jy / beam>}

[e-koch]

Merging. Tests failures are unrelated.