Add purely-parallel measurement interface

[jbosboom]

During the 6/4/2014 OpenTuner meeting, we discussed test parallelism. The SMB demo trials are embarrassingly parallel, with no sequential step needed. Jason mentioned we should add a new measurement interface class for this use case with just a parallel run method rather than the parallel compile/sequential run split.

[jbosboom]

It was easy enough to convert the SMB tuner to parallel -- rename run to compile and change its signature, add run_precompiled that just returns the compile result, and add a new run that's just pass.

While N runs will start in parallel, OpenTuner waits for all of them to finish before starting a new batch. On my dual-core Surface, I can just oversubscribe trials to avoid wasting too much CPU time, but it seems like on a large machine like Lanka many cycles will be wasted waiting for the longest-running trial. It's probably hard to change this behavior now, but it'd be nice to have the full N trials running at all times.