Introduces local_t, a type which is like atomic_t, but the
operations are only atomic from a single CPU's point of view
(ie. atomic against interrupts and softirqs). Some architectures
(eg. i386) can implement these very efficiently.

There are special operations for the case of a local operation on a
per-cpu variable (which is common), which some architectures can
implement efficiently (eg. IA64 keeps all per-cpu variables mapped
at the same address, so no address arithmetic is needed).

The generic implementation given here simply uses atomics on 32-bit
archs, three variables on others.  x86 is already implemented specially.