Reduce SpatialQueryPipeline overhead

[Jondolf]

Objective

The SpatialQueryPipeline currently has some egregious overhead for scenes with a large number of entities. It allocates an EntityHashMap with all colliders (Arced) and some other data, from scratch, every frame. Even worse, it allocates QBVH nodes for all entities up to the largest collider Entity index. For a scene with 999,999 non-physics entities and one collider, it seems to allocate a million QBVH nodes every frame. This produces profiles like this:

Yikes!

Solution

Just store QBVH proxies in a Vec, and use those vec indices rather than Entity indices for QBVH building. This also lets us replace a bunch of hashmap lookups with vector indexing, get rid of some Entity conversions, and avoid table scans for Added<Collider>.

This is still kind of a duct-tape fix. We should ideally have separate BVHs for dynamic or kinematic and static or sleeping bodies, along with incremental updates. It would also be good if we didn't have to store all the duplicate collider data and just queried the ECS directly, assuming the query overhead is negligible. This will be worked on in the future.

Still, this seems to be much better!