Long-running activity in scheduler_run() blocks network, creating information leaks

This is most easily observed when one or two guards are manually configured and all guards becomes unreachable because either they are either

a) down

b) enforcing circuit DDOS rate limiting

One also observes brief sharp traffic forward drop-offs when new circuits are being created rapidly but otherwise under nominal operation.

Log message accompanying the problem:

Failed to find node for hop 0 of our path. Discarding this circuit.

Attached Vidalia bandwidth graphs show one each of (a) and (b) above.

Easy to reproduce.