Testbench: Parallelize thread execution/initialization

I have a chunk of IR that's slowwwwwwwwww to compile and a 40-some core machine. Having to wait 40 * a long time for the threads to start up is painful. Instead, we should JIT-compile in parallel, but without affecting the measured execution time. IOW, the testbench should spawn N threads, have them run init, then have them run actual test stuff.