Support non-NaN nodata values in `mosaic`

[gjoseph92]

From microsoft/PlanetaryComputerExamples#108 (comment).

Support a nodata=np.nan kwarg to stackstac.mosaic, so you can specify a different nodata value if that's what you're using.

The larger problem here is that we're not storing the nodata value in xarray metadata anywhere—just using the xarray convention that NaN is the standard nodata value. But if you set fill_value, that's no longer the case.

Broadly related to #50: rioxarray has a convention for storing nodata in attributes, which we could follow, it's just convoluted CF-style. A single nodata coordinate might be more sensible to xarray users. (The real answer is native missing-value support in xarray though pydata/xarray#4143 pydata/xarray#3955 pydata/xarray#1194.)

If we had this, mosiac could automatically use the nodata value stored in the data, not requiring you to keep track of it.

cc @TomAugspurger

[TomAugspurger]

Thanks I was debating whether or not to open this :) I'm probably OK with just a keyword.

I agree that this really needs to be solved upstream in xarray / NumPy.

[gjoseph92]

I think a keyword is our best stopgap for now, plus maybe storing the metadata. Do you know if there's been any more activity around missing-value support for xarray/NumPy? I haven't kept track for a while.