Optimize collision traversal and add arena memory modes

[VirxEC]

Summary

• Replace the binary BVH with BVH4 and optimize collision/raycast traversal.
• Reduce hot-path work through packed broadphase keys, cached wheel contact velocity, filtered ray candidates, and more compact collision layouts.
• Add Heavy, Balanced, and Light arena memory modes:
  • Heavy prioritizes throughput.
  • Balanced cuts per-arena memory with ~1–3% overhead.
  • Light effectively disables grid partitioning for minimum memory.
• Clean up and restrict internal physics/math APIs.

Performance

stress_v3 throughput improves from 577k TPS to 660k TPS (~14%).

Memory modes reduce marginal active-arena memory from roughly 915–946 KiB in Heavy to:

• Balanced: 108–123 KiB (same as the old light mode)
• Light: 66–91 KiB