Avoid floating point math when parsing coordinates from XML attributes

This can cause loss of precision even on 64 bit architectures, and
apparently even more so on 32 bit. Mainly matters for unit test as
we don't use the OSM XML format elsewhere anyway.