Handle refcount(9) wraparound.

Attempt to mitigate the security risks around refcount overflows by
introducing a "saturated" state for the counter. Once a counter reaches
INT_MAX+1, subsequent acquire and release operations will blindly set
the counter value to INT_MAX + INT_MAX/2, ensuring that the protected
resource will not be freed; instead, it will merely be leaked.

The approach introduces a small race: if a refcount value reaches
INT_MAX+1, a subsequent release will cause the releasing thread to set
the counter to the saturation value after performing the decrement. If
in the intervening window INT_MAX refcount releases are performed by a
different thread, a use-after-free is possible. This is very difficult
to trigger in practice, and any situation where it could be triggered
would likely be vulnerable to reference count wraparound problems
to begin with. An alternative would be to use atomic_cmpset to acquire
and release references, but this would introduce a larger performance
penalty, particularly when the counter is contended.

Note that refcount_acquire_checked(9) maintains its previous behaviour;
code which must accurately track references should use it instead of
refcount_acquire(9).

Reviewed by: kib, mjg
MFC after: 3 weeks
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D21089