[DON'T MERGE] Disable GC on scheduler at benchmark start

[randerzander]

Based on, dask/distributed#4987, ran SF1k w/ GC disabled on the scheduler process, and observed total benchmark runtime improvement of about 5s. This was consistent across many runs.

Results from larger scale tests are pending.

[jakirkham]

Well the other interesting thing from that thread was pinning the Scheduler to a single CPU. Admittedly that's probably a little hacky (ok for the exploration Gabe is doing). If that is important longer term, we can investigate other ways to set that.

[gjoseph92]

@jakirkham worth noting that I only did that so the measured CPU times (particularly idle time) were easier to understand on a multi-core machine. The disabling-GC performance boost seemed the same without it. I suppose it could make a difference with memory access performance if the kernel is moving the process around between physical cores, but I don't think we're at the point of worrying about that yet.