Refactor p2p.PeerTracker

[StephenButtolph]

Why this should be merged

• Refactors the peer tracker to perform all operations in O(1) time.
• Removes the peerInfo struct
• Removes unintuitive handling of TrackBandwidth(..., 0) by explicitly requiring the caller to choose to either call RegisterResponse or RegisterFailure
  • This now also registers the 0 value into the global metrics averager.
• Improves logging inside of GetAnyPeer (now called SelectPeer).
• Moves removal of the peer from the bandwidth into RegisterRequest rather than SelectPeer so that SelectPeer can be read-only.
  • This required a slight restructuring of the network client to hold the lock between the call to SelectPeer and RegisterRequest. This was done just to maintain the prior behavior... It could be reverted if we wanted to allow potentially sending multiple requests to the highest bandwidth providing peer.
• Fixes a minor bug in how bandwidth is calculated. epsilon is documented to avoid division by 0... But it was actually being used to prevent marking a node as unresponsive if it responded with a 0 byte message. Previously we should have been using epsilon in both these cases... Now we only need to use it to prevent the division by 0.

How this works

For the most part - this is a pure refactor.

How this was tested

• Existing CI