Fast user-space mutex implementation, allowing user space to do all
of the normal handling, with a minimal fallback to kernel space for
when there is lock contention.

The kernel space implementation does not keep any per-lock data
structures, but instead does a fast hash on the physical page and offset
of the user-space lock when contended.  Thus no build/teardown costs, or
any scalability costs wrt metadata.

Updated syscall numbers for 2.5.6, and changed FUTEX_UP/DOWN definitions
to be more logical for future expansions (eg.  r/w).