ReSTIR merged GBuffer pass

[kvark]

Experiment branched off #161

The idea is to only fetch GBuffers once at init of the workgroup. We can place that data in workgroup memory and re-use spatially. This would reduce the amount of VRAM traffic (and latency).
In practice, it turned out to be significantly slower:

I suspect this could be due to:

• the driver being better at occupancy when shaders are smaller. NSight confirms this to some extent. It's still not entirely straightforward, since NVidia can have variable register occupancy during the shader execution.
• separate gbuffer pass is more local, it doesn't re-shuffle the groups into clusters
• gbuffer pass able to mix the latency of VRAM access with RT core utilization, while the merged pass becomes more blocked on RTCore

Update

Can confirm this is due to locality as the biggest factor. Here is a run with group shuffling disabled. It's much faster.