Remove hard-coded scratch registers (r9 and ip) from the code generator in favor
of using the `UseScratchRegisterScope` utility. And as a result, we can free the
r9 register for the allocator to use.

Note that the code generator now has to cope with a single scratch register (ip)
instead of two (ip + r9). Therefore the code sequences emitted by moves aren't
as optimized as they used to be. For instance, we now use a scratch S register
in places where we could use r9. We can optimize them further if we want but
running benchmarks showed no impact so keeping the code simpler was deemed
better for the time being.

Bug: v8:6553
Change-Id: I7fcf244cb1b6578564b503619a041006eaf74626
Reviewed-on: https://chromium-review.googlesource.com/895461
Commit-Queue: Pierre Langlois <pierre.langlois@arm.com>
Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Reviewed-by: Bill Budge <bbudge@chromium.org>
Cr-Commit-Position: refs/heads/master@{#51080}