Repairs can stall on a dead fork

[carllin]

Problem

Repairs have a limited quota of requests per iteration, which can be exhausted on a dead fork.

For example imagine a fork structure that looks something like:

          / ----3 -----4 (minor fork)
0 -> 1 -> 2
          \-----------------5-----6-----7---...---1000 (major fork)

Now if slot 3 has more than MAX_REPAIR_LENGTH number of shreds (very probable after a partition, it's all ticks), then there's a chance another validator trying to repair could be missing more than MAX_REPAIR_LENGTH shreds in slot 3. Because repair searches for missing slots starting from the root, in order, then it'll saturate all its repair bandwidth at slot 3, and never try to repair the main fork

Proposed Solution

Weight forks based on number of validators that have

1. > 34% confirmed a fork
2. If not 1), then number of validators that have completed the fork