This PR introduces a parallelism limit during garbage collection. This
is used to avoid large resource usage spikes if garbage collecting many
allocations at once.