Double QBFT timers for lead rounds

[corverroos]

🎯 Problem to be solved

We currently reset QBFT round timers on receipt of any justified pre-prepare even for the current/active round. This results in non-leader increasing their round durations (up to double that of leader) which results in the round 1 leader sometimes getting very out of sync with the rest (since leaders always reset at the start of the round which is an effective noop). This causes sporadic consensus timeouts in the case where no redundant peers are online.

🛠️ Proposed solution

Refactor round timer to double the original round duration irrespective of when then "reset" is called.

[corverroos]

These logs show how the leader get out of sync with others due to non-leaders resetting round timer almost at the end of round 1

[corverroos]

So the idea behind reset the timer on JustifiedPrePrepare is: "short rounds if leader is down, extended rounds if leader is online"

the problem is the round extention isn't aligned. A different approach is to say, if I get a PrePrepare, I double the round duration. So instead of resetting, which subtracts the remaining round duration and adds a full round duration, one just add a round duration. That should mitigate network latency affects.

Previous

 |----1----|
     PP
      |----1----|

Suggested

 |----1----|
     PP
           |----1----|

So as long as the PrePrepare is received anywhere in the round, all nodes "extend" their rounds in the same way, so the rounds remain aligned to their relative start times.

Node1:    |----1----|
          PP
          |----1----|   

Node2:        |----1----|
                       PP
                       |----1----|

Node3:  |----1----|
             PP
             |----1----|

vs

Node1:    |----1----|
          PP
                    |----1----|   

Node2:        |----1----|
                       PP
                        |----1----|

Node3:  |----1----|
             PP
                  |----1----|

This means the round extension doesn't increase "out-of-syncness". Unless a node receives the PrePrepare in the next round. Then it does increase "out of syncness", but that is also preset in existing strategy and is the trade-off of extended rounds in general.

Node1:    |----1----|
          PP
                    |----1----|   

Node2:        |----1----|
                          PP
                        |----2----|

Node3:  |----1----|
             PP
                  |----1----|

One way to mitigate this is to omit large values inside the pre-prepare and by sending that as a separate second message. But this is out of scope of this ticket.