vk/radv: DRI_PRIME optimisations
These changes are based on the recent work done on radeonsi:
The end goal is to limit the performance hit caused by rendering on the dGPU and presenting on the iGPU.
In my limited tests, these changes remove almost all DRI_PRIME overhead and I get ± the same performance in DRI_PRIME mode vs when the dGPU is connected to the screen.
Marked as draft because while this works it's lightly tested and probably requires refactoring (eg: which bits should live in radv and which should live in wsi).