feat: parallel incremental BVH + alternative sweep toi implementation

[sebcrozet]

• Implement a parallel version of the incremental BVH updates (got some help from Claude code to fix potential unsoundness/aliasing issues in the parallel code).
• Implement box2d-style Time-of-Impact calculation.
• Always keep 4-lanes simd enabled (with automatic scalar fallback on unsupported platforms).
• Add supports for optional 8-lanes SIMD (mostly used in rapier, but most of the common simd declarations live in parry).