Profile shuffle w/pending changes

[jakirkham]

Adding PR ( dask/distributed#4270 ), PR ( dask/distributed#4365 ), and PR ( dask/distributed#4375 ) together locally. We can get transition down to ~8% compare this to the ~15% we had originally ( #25 ). So transition would be ~2x faster with all of those PRs merged. It's worth noting this includes ~2% of the time just canceling tasks after each iteration.

Maybe this can be improved more by annotating all Scheduler attributes using this strategy ( dask/distributed#4365 (comment) )? It's worth noting that we are spending ~15% of the time in TCP communication. There's also miscellaneous asyncio, Tornado, logging, etc. that are scattered around. It's a little difficult to understand how much time is eaten up by this as it can be mixed with other stuff we care about.