Use Priority Flood algorithm to fill depressions

[groutr]

Fixes #239

Runtime is about 7.5min on a (12150, 9622) float32 raster which is acceptable.

This variant of the algorithm (as presented in the paper linked in the issue) uses a plain queue to avoid some of the overhead of using the priority queue. Since numba doesn't have a plain queue, I use a list that gets flushed periodically.
A critical part of making this fast was using numba's typedlist which still seems to be experimental according to the docs.

I'm open to feedback.

[groutr]

@mdbartos I've implemented most of the variants of the priority-flood algorithm described in this paper into a priority-flood module on my machine. There are methods for finding the basins, filling depressions, computing flow direction, and resolving flats (though I had to slightly deviate from the algorithm in the paper to make in work for test DEM). Do you want these other methods in pysheds too?

[groutr]

@mdbartos I don't think the nodata_out parameter is needed any longer in fill_depressions. How do you feel about deprecating it?

[mdbartos]

• Sure, I'm fine with deprecating it.
• I'm interested in the other methods, but I will pull in this PR first. Feel free to submit another PR with these methods included if you think there's value.

Thanks again and awesome work,
MDB