If a build has a systematic failure, we currently just let nodepool
run, looping builds until we hit the overall job timeout (1.5 hours).
This adds a count of output log files; if we see three failed builds,
then assume the problem won't get better and fail early.
Change-Id: Id7e163b4937dd57cc8afbf72ed795f73b46a05b1
899fb6e368