Depth buffer injection improvements

[David Dubois]

Some things to improve the depth buffer injection:

• Generate a mip chain for the normal target as well
• Can we use FP16 when generating the depth buffer mip chain?
• One pass to inject into all cascades. The VS reads the point's position and normal and passes it to the GS. The GS tests that position against each cascade, and emits a primitive for each cascade that passes. the FS adds the SH normal as usual

[David Dubois]

FP16 depth is fine?

One pass to inject cascades took my total time from like 4 ms to 3 ms. It's good, but it could be better. Using mip 1 of the mip chain got me more gains

My occupancy is trash. Nsight says I'm blocked on passing attributes from stage to stage. I've converted to fp16 but it didn't make a huge difference. The whole thing isn't that slow, but I feel like it should be better