Implemented 2.5D culling. (Basic idea: after computing min/max depth for the tile, slice it into 32 ranges, build a bitmask of depths that are present. Build the same for lights. Only accept lights that intersect the depth mask.)

From these A/B shots (false positive heatmap), you can see that it does help. A little. But the mask for the light is very crude, just uses the min/max in view space, doesn't account for the bounding sphere.

Okay, remapping with:
`log2(float(count) + 1.0) / 2.0`

This gives blue at 3, green at 15, red at 63, and white at 255:

Heatmap view, with the light count in the scene raised to 256.

There's two culling tests, and a light is only marked visible if it passes both:
- inside the planes of the subfrustum
- inside the viewspace AABB of the subfrustum

AABB only is decent, but falls flat on edges. Frustum only is _awful_. The two together do a pretty good job, though.