Shrink stack size for unused threads

[btrask]

We use libco for fibers which allocate their own stacks. That means we basically don't use the default stack for the main thread and the worker threads we spawn. I don't think there's a way to shrink the main thread once the process is already running, but we could cap the worker threads before we spawn them.

See also pthread_attr_setstacksize, PTHREAD_STACK_MIN

Not sure about portability. Could be POSIX-only for now.

Edit: This would save a little bit of address space but I guess the main advantage would be fragmentation? If the default stack space is 8MB and we spawn 16 worker threads, that's almost 128MB we're wasting (but the kernel VM means it's not as bad as it sounds).

Edit: It's a bigger deal on 32-bit.