Quantity.__array_function__ follow up for Quantity

EDIT: non-Quantity discussion of __array_function__ is in #8610

In the non-yet-merged but ready-for-review #8808, support for numpy’s __array_function__ is introduced. For a number of functions, a decision still needs to be made, mostly on whether to stick exactly to the numpy docstring or to make adjustments to parameter defaults that make the function more useful for Quantity

• np.isclose, np.allclose: these were never all that useful since internally they calculate abs(a - b) < atol + rtol * abs(a) and atol is given as a small number, which is interpreted as dimensionless. For Quantity, we could either set it to 0 or assume the units of a, but in either case it would become inconsistent with the docstring that users actually see. (Decision: units of a - #9463)
• np.broadcast_arrays and np.broadcast_to by default return plain arrays, needing subok=True to work for Quantity. We could change the default if a Quantity is present. Again, the possible issue is that the docstring will no longer be correct. (Decision: leave as is, since old code could cease working; #9463)

Furthermore, a few top-level functions need an actual implementation and/or better test cases:

• np.histogram2d and np.histogramdd (see #8900)
• np.piecewise (see #8884)
• np.apply_along_axis and np.apply_over_axis (see #8900)
• The setops functions such as np.unique might as well be supported (see #8900)

And finally we should also cover the numpy submodules that have routines not exported to the top level. In particular:

• np.fft (#9464)
• np.linalg (see issue #8622; PR #9494)
• np.polynomial (has no functions that can be overridden).

EDIT: And it would probably be good to add tests for Quantity subclasses:

• LogQuantity such as Magnitude
• Angles such as Longitude

[bsipocz]

@mhvk - I’m removing the milestone from here as there are no more open PRs ready for 4.0 that are related.

[mhvk]

Since this is mostly done, closing - leftovers at #10575.