FIX ignore masked voxels when averaging with nanmean=True

[TomDLT]

When averaging over voxels to produce a flatmap, NaNs can be ignored using nanmean=True.
However, if the data is masked (e.g. with a "thin" mask), masked voxels are not ignored like NaNs.

Example with data filled with 1s and NaNs:
"thick" mask: correct behavior, averaging 1s into 1s, and NaNs into NaNs.

"thin" mask: incorrect behavior, averaging 1s into value in [0, 1].

The fix is to use masked_data.filled(), which fills masked numpy arrays with the value in fill_vallue.

import cortex
import numpy as np
from matplotlib import pyplot as plt

subject = 'S1'
transform = 'fullhead'

for type_ in ["thin", "thick"]:
    # create dataset with volume from 0 to n_voxels
    mask = cortex.db.get_mask(subject, transform, type=type_)
    vol = cortex.Volume(np.ones(mask.sum()), subject, transform, cmap="viridis", vmin=0, vmax=1)

    # set 50% of the values in the dataset to NaN
    vol.data[np.random.rand(*vol.data.shape) > 0.5] = np.nan

    _ = cortex.quickshow(vol, nanmean=True, with_curvature=False)
    plt.show()

[TomDLT]

Actually, this fix is incomplete, because the same problem arises with nanmean=False:
thick mask:

thin mask:

[TomDLT]

Full fix
Before:

After:

[mvdoc]

Thanks for this! the docs/tests are failing, and It looks like it might be failing when a non-masked numpy array is passed

[mvdoc]

awesome, thank you!