Improve DRM use of tmpfs for GEM buffers

Currently, DRM uses shmem/tmpfs to support GEM buffers. This mechanism provides lots of benefits: virtual memory subsystem control over the backing pages (without resorting to userptr), and a backing 'struct file' needed to properly implement dmabuf mmap support for cached buffers. However, for some of the memory management mechanisms being employed for ARM SoC support, the driver will need to have more control and flexibility around page allocation. Supporting allocation schemes like CMA for drivers that need contiguous buffers is not so straightforward with the current arrangement. We will explore enhancements to this mechanism to make it both more flexible and, ultimately faster for the ARM environment.