Move registerDelay in MonadConc?

[NicolasT]

Currently, registerDelay is implemented in function of what's provided by MonadConc, forking a thread. This works, but the registerDelay implementation from GHC/stm is quite different: it schedules something in the runtime's timer manager to get things done at some future point in time, removing the need for a thread.

Would it make sense to have registerDelay in MonadConc so the IO instance can use the one from stm, and Program the current implementation (using a thread)? Operationally, both should be the same (otherwise the registerDelay in concurrency wouldn't be interchangable with the one from stm in the first place...), so tests using DejaFu should still find bugs, even when in IO another implementation is used.