Implement new WorkGraph push/pop

[ibaned]

debugging #988 revealed less-than-ideal aspects of the current WorkGraph push/pop implementation. @hcedwar suggested a new design, this issue is to implement that new design and compare to the current one in terms of performance and robustness.

[hcedwar]

New algorithm has only a single state for push/pop of the ready queue and uses the ready queue begin/end indices as performance hints. This entirely avoids the problems with updating a split-state (#988). Also replaces a compare-and-swap loop with a simple atomic increment.

[hcedwar]

Pop function should not spin waiting for work. Instead return either (a) a valid work index, (b) value indicating that all work is done, or (c) value indicating that no work index is currently available. The back-end specific scheduler should be handling condition (c), not the pop function.