On relay close, clean first, reply later

[karalabe]

Currently the relay service immediately acknowledges a client connection closure request. This however leads to a race condition where clients think they've been cleaned up, but in reality not yet.

This becomes obvious during testing (iris-go: req/rep), that one test influences another. If they are run individually, all is ok. If multiple one after the other, new requests get assigned to stale old connections.