Policy change: don't retry benchmarks that time-out

[rrnewton]

Benchmarks that timeout are by definition the biggest time consumers... retrying those can be really bad on the bottom line for job runtime.

One option is to not retry on time-out rather than crash. Retrying is more for testing programs that are still buggy and have some chance of crashing.

However, nondetermainistic bugs can also manifest themselves as a livelock. That's one argument for retrying but perhaps having an exponential decay of the timeout, so that the benchmark cannot consume too much total time via retries.