For now this can only sleep in increments of the scheduler quantum
(currently 50ms). It also uses a somewhat ineffecient way of tracking
the sleeping threads - it will scale linearly with the number of
sleeping threads.
This allows us to easily track the physical memory so it
can be freed when the thread exits. It also simplifies the page fault
handler as it just needs to check regular mappings to find a user stack.
Instead of passing an enum with the capability when creating it, relying
on polymorphism and a template struct tag to determine the object type
at runtime.
This is cleaner and avoids errors where we pass the wrong capability
type with the cap and do a bad cast at runtime.
