Yet more improvements to the unadvertise test

[jcorbin]

• drop the 2ms (re)advertise interval: it was never needed, and causes flaps once we start injecting more delay by deferring peer.connectTo
• tighten then connection removal expectation fence to be all exit connections (both in and out), and not just any in-connections
• simplify unadvertise tail flow, turns out we didn't need a counting fence
• trivial improvement to declined assertions for make clearer fail output

r @rf @raynos @kriskowal