Fix distributed_work segfault when local work generation fails

[guilhermelawless]

In start_work () :
else if (!this_l->cancelled && !this_l->completed)

Was not setting any of these atomics to true, so the same callback would possibly be called in cancel_once. Not normally an issue, but can happen and came up with Serg's epoch storm test. Instead of making a new atomic, joined both into a single finished atomic, and added a work_generation_status enum.

Also added a check if (callback) on every call.