We could have limped along without this for v10, which was my intention
when I annotated the bug in commit 76a3df6e.  But consensus is that it's
better to fix it now and take the cost of a post-beta1 initdb (which is
needed because these node types are stored in pg_class.relpartbound).

Since we're forcing initdb anyway, take the opportunity to make the node
type identification strings match the node struct names, instead of being
randomly different from them.

Discussion: https://postgr.es/m/E1dFBEX-0004wt-8t@gemulon.postgresql.org