Meshlet tracking issue

[JMS55]

Runtime performance:

• 🔥 (Probably not needed once we have BVH-based persistent culling, but super easy and should be a nice win): Rather than a binary search in fill_cluster_buffers with one thread per cluster, spawn 1 workgroup per instance, have it reserve a slice of the cluster buffers, and then have its threads loop writing elements to the instance and meshlet ID buffers.
• 🔥 Cleanup resource_manager.rs and visibility_buffer_raster_node.rs, improve CPU performance, and add more GPU debug group labels
• Fix all occlusion culling bugs
• Persistent culling

Runtime performance (low priority):

• Write a heuristic for software vs hardware raster selection
• Upload instance uniforms only for changed entities (and see if we can shrink the size too)
• Single set of passes to render all views at once
• Instance culling
• Compute-based material shading
• Variable-rate shading (compute shader based)
• LOD bias for shadow views

DAG building:

• Manual vertex locking
• Larger meshlet groups
• Vertex attribute-aware simplification
• Multithread parts of MeshletMesh::from_mesh() maybe? (could use rayon)

Compression:

• Meshlet triangle compression
• Streaming
• Disk-oriented compression

Tooling:

• Debug stats + visualizations
• Asset processor for mesh -> meshlet mesh conversion
• Docs