Hello Matthias,
in address family vpnv4 the next-hop is changed automatically, because it needs to be an address in the global routing table and so the eBGP next-hop learned in the eBGP session in VRF cannot be reused.
This is needed in order to have a working forwarding plane
This does not happen in af ipv4 because the original eBGP next-hop can be retained being part of global routing table on ASBR/PE node, in this case setting next-hop self may solve problems of installation of BGP routes in downstream iBGP peers as they could not know how to reach theoriginal BGP next-hop.
Hope to help
Giuseppe