Ensure the bin_edge version of `SpectralAxis` specification has a consistent wcs

It's not clear to me that the implementation of #645 (and future additions like #661) fully follow the convention that if the user does `spectrum.wcs.pixel_to_world([0,1,2,3,...])` (i.e. integer pixels) they get the bin-edges, while `spectrum.wcs.pixel_to_world([0.5,1.5,2.5,...])` should give the centers. A test of this should be added to ensure this is the case, and fixed if not.

This is a follow-on to #645 based on some of the discussion in #176 and https://github.com/astropy/specutils/pull/645#discussion_r405718570 .

cc @rosteen

Start by reviewing the implementation of #645 and the discussion in #176 and the linked pull-request comment. Add a regression test comparing pixel_to_world results for integer pixels and half-integer pixels, then fix the implementation if the integer inputs do not produce bin edges and the half-integer inputs do not produce centers.

Written by the indexing model from the issue text.