Update `mask` to allow use of native `nodata` value

[robbibt]

Currently the .odc.mask() method uses .where() to set pixels to NaN:
https://github.com/opendatacube/odc-geo/blob/develop/odc/geo/_xr_interop.py#L311-L359

It would be good to allow the function to set pixels to the native nodata value of the data instead.

[Kirill888]

Just a note that this needs to be done carefully to avoid breaking the original "contract" of having a side effect of converting to floating point. Default should remain as is, I feel, but an option should exist that allows keeping original dtype on the output.

[robbibt]

Yep, makes sense - right now there's not any way to configure nodata, so at the very least it would be great to be able to pass in a custom value rather than defaulting to NaN.

[Kirill888]

Yep, makes sense - right now there's not any way to configure nodata, so at the very least it would be great to be able to pass in a custom value rather than defaulting to NaN.

@robbibt that's by design, because output is always float, there is no need for nodata marker, since NaN is THE nodata for floating point data. Using anything but NaN as a fill value for float output is odd and confusing.

The real parameter here is: do you want floating point data with NaNs on output OR do you want to keep the original dtype and original nodata of the input on the output. Currently there is no choice, you always get float with NaN, we can add some parameter, maybe mode: Literal["float"] | Literal["original"] | Literal["auto"] = "float". One could also add nodata parameter I guess, but then you'd have to deal with the case where output nodata is different from source nodata and will need to be remapped.

Looking at the code it seems that it's missing the step that converts original nodata into NaN, so you end up with NaN and also potentially with whatever original nodata marker was as well, basically this method is broken for non-float inputs in it's current state.