After yet more review comments from several people, here is a reworked
set of 32-bit MMIO copy patches.  This may even be the final set.

These define the generic __raw_memcpy_toio32 as a weak symbol, which
arches are free to override.  We provide a specialised implementation
for x86_64.

These patches should apply cleanly against current -git, and have been
tested on i386 and x86_64.

The patch series is as follows:

raw_memcpy_io.patch
  Introduce the generic MMIO 32-bit copy routine.

x86_64-memcpy32.patch
  Add memcpy32 routine to x86_64.

arch-specific-raw_memcpy_io.patch
  Get each arch to use generic memcpy_io code, except x86_64, which
  uses memcpy32.

Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>