arch/xtensa: Inline atomics

The xtensa atomics layer was written with hand-coded assembly that had
to be called as functions.  That's needlessly slow, given that the low
level primitives are a two-instruction sequence.  Ideally the compiler
should see this as an inline to permit it to better optimize around
the needed barriers.

There was also a bug with the atomic_cas function, which had a loop
internally instead of returning the old value synchronously on a
failed swap.  That's benign right now because our existing spin lock
does nothing but retry it in a tight loop anyway, but it's incorrect
per spec and would have caused a contention hang with more elaborate
algorithms (for example a spinlock with backoff semantics).

Remove the old implementation and replace with a much smaller inline C
one based on just two assembly primitives.

This patch also contains a little bit of refactoring to address the
scheme has been split out into a separate header for each, and the
ATOMIC_OPERATIONS_CUSTOM kconfig has been renamed to
ATOMIC_OPERATIONS_ARCH to better capture what it means.

Signed-off-by: Andy Ross <andrew.j.ross@intel.com>