unit-aware versions of unsupported functions

[keewis]

See also #6.

Due to a variety of reasons, some functions don't support units (partially or not at all). Since that seems unlikely to change in the near future, we can provide wrapper functions that handle the units before passing them to the appropriate xarray function.

I'm currently planning to add these:

• sel (sel #60)
• drop_sel (implement drop_sel #73)
• loc (implement loc #79)
• reindex / reindex_like (reindex and reindex_like #69)
• interp / interp_like (interp and interp_like #72)
• ffill / bfill (bfill and ffill #78)
• chunk (to convert wrapped numpy to dask) (chunk #83)
• interpolate_na (interpolate_na #82)

[keewis]

I will hold off on interpolate_na because it might be possible to replace the call to numpy.vectorize by broadcasting correctly (which, considering that vectorize is essentially a python loop, might also be faster).

[keewis]

after discussion it seems that supporting interpolate_na is difficult right now: numpy.interp does not support the axis kwarg (which is the reason for vectorize) and scipy does not support duckarrays. I guess we'll need a wrapper function for it.

[keewis]

as for chunk: is chunking after quantify something we should support? rechunk does work properly, but da.from_array will use pint.Quantity as meta.