Fix bug in GCRS<->ICRS conversion for near-earth objects #4221

As has been discussed a fair amount in the astroplan project, as well as on the mailing list (the thread starting with https://mail.scipy.org/pipermail/astropy/2015-August/003907.html), there's a significant bug with the GCRS<->ICRS transformation for coordinates that are close to the earth (seems to be within <~1 AU). For a regression test demonstrating the problem, see the first few commits in https://github.com/eteq/astropy/tree/fix-gcrs-trans

I'm pretty sure I understand the source of this (basically, it's due to an incomplete workaround to some of ERFA's limitations), and hence am assigning it to myself. Making this issue mainly so that we can try to get it fixed before a final v1.1 release.

An item to consider, though: once fixed, should this be included in the 1.0.x branch? While it is a bug in that the transforms aren't right, it is technically an API change in case someone worked around it already... (@astrofrog and/or @embray)?