Wrong behaviour of copy() method with new_array from np.ndarray with NaNs

[rhugonnet]

The copy method expects a np.ndarray as input for the new_array argument. However, if that array contains NaNs, the Raster object created will have as .data a np.ma.masked_array which contains unmasked NaNs.

This results in a lot of issues: applying np.ma functions, writing a Raster to disk in rasterio, propagating nodata when applying arithmetic functions to Raster objects, etc... Strangely, some systems support the writing of these unmasked NaNs to disk in TIFF, while it result in a break down on others.

We should probably check the input type in copy, or upstream in from_array i.e. in Raster.__init__ if the issue is global to these routines, in order to create a masked_array that masks the NaNs automatically.