Fix encoding of masked dask arrays

[jsignell]

• Closes Reading data after saving data from masked arrays results in different numbers #9374
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

I really tried to write a simpler test, but I couldn't quite get it to work.

[dcherian]

We don't support np.ma:

xarray/xarray/core/variable.py

Lines 295 to 301 in 2742e22

	if isinstance(data, np.ma.MaskedArray):
	mask = np.ma.getmaskarray(data)
	if mask.any():
	_dtype, fill_value = dtypes.maybe_promote(data.dtype)
	data = duck_array_ops.where_method(data, ~mask, fill_value)
	else:
	data = np.asarray(data)

though we'd like to support marray ASAP: #10574

In this case, i think np.ma is sneaking through because we don't check the dask array's meta before wrapping in Variable

[jsignell]

We don't support np.ma

Oops should have checked that first I guess :)

[dcherian]

should we use filled for the np.ma path too? Or is it inplace and weird to do that?

[jsignell]

filled does seem like it was clearly designed for this purpose and it doesn't seem like it's inplace. I think you are right. That would be clearer

[jsignell]

Well in retrospect there is some type promotion going on in where_method so maybe the dask version should also just use that.