A memory ordering barrier is needed in crit_exit() ensure that td_pri

A memory ordering barrier is needed in crit_exit() ensure that td_pri
is adjusted before gd_reqflags is tested.

Reorder the tests in crit_exit() as a branch prediction optimization.
Check gd_reqflags first, since it is likely to be NULL, and td_pri second,
since thd td_pri check may oscillate depending on the critical section's
nesting level.

Load the globaldata pointer before adjusting td_pri in crit_exit_quick()
as an instruction ordering optimization.