Without this cap, unlucky scheduling and/or details of what pipeline
stages perform best on the host machine could cause many gigabytes of
MIR to be stuck in the queue. At a certain point, pause the main thread
until some of the functions in flight have been processed.