Normalization of remote Coordinated Main Relations

[danielhers]

The conversion from json checks if the edge has a CMR category and then sets a CMR attrib: https://github.com/danielhers/ucca/blob/master/ucca/convert.py#L968
Then the normalization finds this attrib, splits the relation and destroys the node: https://github.com/danielhers/ucca/blob/master/ucca/normalization.py#L260

A remote CMR is expanded such that the second scene points only to the first conjunct.
This way the last, remote-only scene is dropped, and the result is not logically equivalent.
It will have a remote which is a CMR that is half-expanded, and includes only the first conjunct (what was originally the first Center).
Otherwise, if we expanded the full Cartesian product, we would get a scene containing only remotes, which is possible - we should mark it as an implicit scene, but we have never had something like that yet.
CMR is not a category we're currently supporting in the final XMLs, and it would be weird to include it only in those cases.

Perhaps CMRs should not be split, but rather be treated just like regular P/S (at least in these cases). If we turn it into LF we could split them, but as it is, we should not split CMRs.