v3d, panfrost, asahi: Do not allocate renderonly buffers for PIPE_BIND_SHARED requests.
As far as I understand PIPE_BIND_SHARED means that buffer can be shared between processes using dma-buf fd (or flink name).
Such buffers will not be used by display controllers (otherwise PIPE_BIND_SCANOUT must be used additionally).
Allocating them using renderonly means to use DRM_IOCTL_MODE_CREATE_DUMB ioctl on display controller node, which will use CMA for most of the platforms.
Moreover renderonly will not select best modifier for such allocation (usually only LINEAR is supported, only high-end platforms has compressed/tiled scan-out support).
freedreno, vc4, lima, etnaviv has already such behavior (but modifier is forced to linear for some of them for unknown to me reason).
Implementing this requires !12370 (merged) merged, otherwise we will face the issue #5225 (closed)