- queue_insert before incrementing s->waiting
- Decrement t->wait only if we might execute t
  i.e. it passes the mask and skip checks
- Remove unused non-blocking queue_gettask operation
- atomic.h atomic_load adds fences, avoids tearing across cache lines
- t->weight compile-time flag disables dependence on previous runtime
- scheduler_start compile-time flag removes race with scheduler_done