topk and argtopk

[Huite]

• Closes Generalized (arg)min, (arg)max: add nsmallest, nlargest, arg_nsmallest, arg_nlargest #10075
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

I've made a start with implementations of topk and argtopk.
This'll work uncontroversially for DataArrays with and without NaNs, although I'm getting mostly stuck on what skipna=True should entail.

There's a number of choices to make, however, which are probably best illustrated with this PR.

• Numpy does not provide a topk implementation, dask does, as @dcherian helpfully pointed out. I've borrowed some parts of the dask implementation and put it in nputils.py. The question that arises here for me: I guess this would work with cupy too, but I don't quite oversee what's the best way to integrate it? Because dask provides topk of its own, it appears a bit exceptional.
• Another question: bottleneck also provides partition and argpartition. This version works differs from numpy.partition in its handling of NaNs, however. Using numpy's partition has the benefit that it's consistent with dask.
• In terms of API: topk feels mostly similar to quantile, since it shortens, but doesn't reduce the dimension entirely. argmin also supports an axis argument next to dim (though exclusive) -- is the axis argument desirable?
• The code in variable.py borrows somewhat from quantile, since the result has an axis with dim k size instead of len(q). Unlike quantile, dask's topk and argtopk do not support tuple arguments for the axis (although topk accepts it and produces an unexpected result), so part of the stacking and unraveling functionality of _unravel_argminmax is required. I've currently duplicated the relevant lines to keep changes clearly visible.
• I'm insufficiently familiar with apply_ufuncs to judge whether dask="allowed" works gracefully with the dask topk and argtopk functions, my guess is that it should.
• quantile returns a result with a new dimension and coordinate called quantile, I've mimicked this and topk and argtopk return a result with a new topk or argtopk dimension respectively. I was thinking no labels are required for topk, but since both positive k values (for largest) and negative k values (for smallest) are possible, it's probably smart to return labels range(0, k) and range(-k, 0)?
• There's also idxmin for the coordinate labels, and I suppose idxtopk would make sense too?
• skipna=False is giving me some headaches. A (naive) implementation as in this PR is assymetric. Numpy partition (and thus the dask version too) sorts NaNs towards the end of the array, such that k > 0 will return NaNs, but k < 0 will not. For the testing, I figured da.topk(k=-1, skipna=False) should equal da.min(skipna=False) and da.topk(k=1, skipna=False), should equal da.max(), but this isn't the case. k=1 will return a NaN value since numpy partition moves the NaN to the end; k=-1 will not. I currently gravitate towards accepting this assymetry, since e.g. np.sort will also move NaNs to the back and it feels forced to fetch NaNs for k=-1 to match .min(skipna=False) . On the other hand, Python's sorted behaves differently, according to IEEE 754 NaNs are not orderable ... and I reckon you'd mostly use skipna=False when you want to ensure that no NaNs are present?
• The NA-handling is currently in duck_array_ops, maybe it belong in nanops as it resembles _nan_argminmax_object and _nan_minmax_object, but is again slightly different. But I didn't like the circular imports that it seems to require; in duck_array_ops it decides whether to use dask or numpy (via nputils), but the masking of NaNs is required for both.