Many newer overlays use IRenderTextures that are sized to the rendered viewport. This was completely broken, because a single viewport can be rendered on multiple viewports in a single frame.
The end result of this was that in the better case, constant render targets were allocated and freed, which is extremely inefficient. In the worse case, many of these overlays completely failed to Dispose() their render targets, leading to *extremely* swift VRAM OOMs.
This fixes all the overlays to properly cache resources per viewport. This uses new engine functionality, so it requires engine master.
This is still a pretty lousy way to do GPU resource management but, well, anything better needs a render graph, so...
8c16b4580b