Make friend schedulers more robust

[brson]

Right now when you spawn a SingleThreaded task, its scheduler is given a 'friend' scheduler to which it sends work that can't be executed. This creates a situation where threads that are rejecting a lot of work serialize a bunch of tasks through a single other thread which may not be able to run that work in a reasonable time.

Instead of sending work to a single friend it needs to either go back into the general pool, or (better) single-threaded schedulers should add their work queues to the global pool of work queues for stealing.

[brson]

We could band-aid over one of the most common deadlocking problems this causes by making the runtime detect situations where a single-threaded scheduler spawns another single-threaded scheduler (creating a chain of single-threaded friends), and doing the spawn from the friend.

[alexcrichton]

This will be dealt with by #10965 because spawned schedulers will be full-fledged citizens of the sched pool (just as much so as everyone else).

[alexcrichton]

Closed by #10965, we no longer have by-default spawned friend schedulers.