discovery: revamp premature update map

[Roasbeef]

Turns out we need it right now to handle some low latency race
conditions in our integration tests, so we'll opt to simply cap the size
of it to a low amount. We use a basic LRU caching mechainsm.

Fixes #5076

[guggero]

The linter fails, which is also the reason for the Travis failure:

 discovery/gossiper.go:1958:34: Error return value of `d.prematureChannelUpdates.Put` is not checked (errcheck)
				d.prematureChannelUpdates.Put(shortChanID, &cachedNetworkMsg{
				                             ^
discovery/gossiper.go:1967:34: Error return value of `d.prematureChannelUpdates.Put` is not checked (errcheck)
				d.prematureChannelUpdates.Put(shortChanID, &cachedNetworkMsg{

I looked at some of the itest failures:

• bitcoind-notxindex: problem with invoice subscription, fixed in itest: fix two flakes #5905
• bitcoind-etcd: Created an issue: itest flake: etcd: HTLC is causing force close when it shouldn't #5904
• bitcoind-postgres: problem when copying entire DB during one of the tests. ignoring for now
• neutrino: best height de-sync problem, created an issue: itest flake: subsystem height de-sync issue "update announcement is premature" #5903
• windows: problem with the sweep happening too soon, fixed in itest: fix two flakes #5905

[Crypt-iQ]

LGTM, left some minor nits 🍳

[positiveblue]

LGTM 👌