Specialize `Deferred` for `IO` to use `CallbackStack`

[djspiewak]

The get function on Deferred relies on notifying listeners of the fulfillment of a write-once memory slot. This is startlingly similar to the implementation of join in IOFiber, except that IOFiber uses a dramatically more optimized data structure than what we find in Deferred. In theory, it should be possible for us experiment with overriding Concurrent[IO].deferred with a custom implementation (IODeferred?) which in turn uses the CallbackStack data structure from IOFiber itself. We should benchmark whether this is actually faster (it should be!) and by how much.

[armanbilge]

Is there a reason we can't / shouldn't move CallbackStack into cats-effect-kernel and use it in the default Deferred implementation?

[djspiewak]

It increases the public API unnecessarily. There isn't really any benefit to it in most cases, anyway, since the construction always indirects through the typeclass anyway, so we can ensure this always applies for IO.