Alloc GEM buffers backed by noncoherent memory on SoCs where it is
actually faster than write-combine.

This dramatically speeds up software rendering on these SoCs, even for
tasks where write-combine memory should in theory be faster (e.g. simple
blits).

v3: The option is now selected per-SoC instead of being a module
    parameter.

v5: - Fix drm_atomic_get_new_plane_state() used to retrieve the old
      state
    - Use custom drm_gem_fb_create()
    - Only check damage clips and sync DMA buffers if non-coherent
      buffers are used
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patchwork.freedesktop.org/patch/msgid/20210523170415.90410-4-paul@crapouillou.net