drm/msm: Use dma_sync_sg_for_device() to flush cache for new buffers

With the conversion of the master device to use the arm iommu
ops we lose the benefit of a cache flush using the dma_map_sg
trick. Plus we end up mapping all the GPU buffers into the master
DMA device which is admittedly sub optimal.

However it did end up mapping all the display buffers which
unintentionally worked. Put the map_sg() functions back into msm_gem.c
to make it intentionally work from now on.

Change-Id: Ic0dedbada1a8723fb47311189a6b7e11969427ee
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>