Previously, if a database request spanned a related object manager, the
first manager encountered would cause a request to the router, and this
would bind all subsequent queries to the same database returned by the
router. Unfortunately, the first router query would be performed using
a read request to the router, resulting in bad routing information being
used if the subsequent query was actually a write.

This change defers the call to the router until the final query is acutally
made.

It includes a small *BACKWARDS INCOMPATIBILITY* on an edge case - see the
release notes for details.

Thanks to Paul Collins (@paulcollinsiii) for the excellent debugging
work and patch.