Parallel driver script and optimizations to reduce system time

[saethlin]

This is the rig I've been using to actually run rustlantis. The objective here is to use as much CPU time as possible, and avoid syscalls where reasonable. The less system time, the more responsive the OS remains while rustlantis is running.

Figuring out where the sysroot is and passing through the rustup proxies is a significant amount of syscall traffic. Avoiding that trims off about a third of all system time.

The other change here is making difftest serial. It is pretty much always more efficient to do parallelism at a higher layer of abstraction, and that's what the script fuzz.sh does; spawning nproc fuzzing instances that are all single-threaded and run completely independently.

[cbeuw]

Avoiding that trims off about a third of all system time.

That's really great!

The other change here is making difftest serial

Could we make parallelism configurable? When manually reducing a program, I often run difftest every time I make an edit. In this case the parallelism across backends are useful

[saethlin]

How about this?