Smoothing with masked data

[dopplershift]

We should double check the behavior of our smoothing functions with masked data and make sure it's what we want. Issues:

• sum(masked_array) is masked if any points are masked
• smooth_gaussian calls gaussian_filter, which is completely ignoring the mask

[dopplershift]

I'm also not entirely sure this fixes anything. This approach (and the current behavior) means that masked values are used within the smoothing operation, then their original location is masked. I would think that users would expect masked values either to propagate or not participate in the operation. We could set them to NaN to make them propagate, but I'm not entirely sure if that's better. Thoughts? @kgoebber ?

Originally posted by @dopplershift in #1935 (comment)

[kgoebber]

I think either way (not using NaN or having them propagate) are both sensible. The issue with some of the smoothing techniques (e.g., smooth_n_point) that can be applied iteratively is that in the propagating NaN scheme you could end up with not a small area impacted. It may make sense not to use NaN values in smoothing, but keep those individual grid points as NaN and not fill them in through smoothing by default.