RPC endpoint `system_peers` stuck at `0` best block number for 1 of 2 peers

To reproduce, `git checkout qdm12/end-to-end/rpc/system_peers`. This branch contains a single commit `wip` adding a check verifying all peers have the same block numbers when querying the `system_peers` endpoint.
This tend to fail if we sleep for a few seconds (`time.Sleep(5 * time.Second)`), where one of the two peers connected is stuck on the best block number `0` (but never both, oddly).
The corresponding RPC handler is `Peers() []common.PeerInfo` but I did not find anything wrong there.

I'm not sure of the severity and depth of this bug, but I cannot 'quick-fix' it in #2470 hence the reason for this issue. Note this was found by adding more testing depth in #2470 but this wasn't asserted before either.