freedreno/a6xx: gmem limits
Fixes a GPU hang I noticed while debugging a crash with 11517 x 12000 png. The crash seems to be triggered by OoM in the renderer process before things even get to the gpu-process (although we might want to figure out a better option than fd_bo_map()
to clear UBWC, to avoid mmap'ing such large buffers, which is problematic with 32b userspace)
But we might as well fix the GPU hang.