Enable multi-coord grouping from xarray

[max-sixty]

I'm working through some around non-trivial groupby from colleagues. In particular, grouping by multiple coordinates seems much harder in xarray than pandas. Flox actually does this really nicely, as per the comment from @dcherian:

As an aside, the API isn't great but this works in flox (I think)

import flox.xarray

flox.xarray.xarray_reduce(da, "labels1", "labels2", func="mean")

Originally posted by @dcherian in #9278 (comment)

Would be great if da.groupby(["labels1", "labels2"]).mean() worked too — is that just a simple translation to flox from the xarray code? (I can probably do it if so). Or is there something more complex going on?

[keewis]

from #6610 (comment) the idea is to eventually support

data.groupby({"labels1": xr.UniqueGrouper(), "labels2": xr.UniqueGrouper()}).mean()

As far as I remember, multi-dimensional groupby is planned but not supported yet. What we'd need to do is remove the condition here:

xarray/xarray/core/dataset.py

Lines 10388 to 10393 in ce5130f

	if len(groupers) > 1:
	raise ValueError("Grouping by multiple variables is not supported yet.")
	elif not groupers:
	raise ValueError("Either `group` or `**groupers` must be provided.")
	for group, grouper in groupers.items():
	rgrouper = ResolvedGrouper(grouper, group, self)

and make sure the loop still works (@dcherian will have more details).

[max-sixty]

from #6610 (comment) the idea is to eventually support

data.groupby({"labels1": xr.UniqueGrouper(), "labels2": xr.UniqueGrouper()}).mean()

Great!

I would have also supported data.groupby(["labels1", "labels2"]) as syntactic sugar (which pandas supports) but I would need to read more to have a confident view...

[dcherian]

Yes I have struggled to fully generalize it.

That said, we could just begin with simply supporting _flox_reduce. This should be a straightforward extrapolation of what's already there.