Skip the heightfield triangles and segments a shape-cast can't reach

[illusion-pasture-program]

Shape-casting against a HeightField shape-casts every triangle (every segment in 2D) of the cells it walks through, including the ones the moving shape can't reach. This resolves the TODO: pre-check using a ray-cast on the Aabbs first? comments: before shape-casting a triangle or segment, cast the moving shape's Aabb against its Aabb, with the same Minkowski-sum ray-cast cast_shapes_heightfield_shape already does against the whole heightfield (and composite shape-casts do on BVH nodes).

I ran into this with CCD on a car chassis over terrain with 12.5 cm cells. Because the terrain's Aabb contains the chassis, every triangle under it (about a thousand) went through GJK every step, and nearly all of them missed.

Speed. New ignored benchmark, heightfield_shape_cast_perf (cargo test --release -p parry3d --test heightfield_shape_cast -- --ignored --nocapture), three runs each:

	before	after
car-sized box riding 20 cm over rolling terrain, one 60 Hz step at 30 m/s	712 µs/cast	59.5 µs/cast
balls falling onto that terrain (nothing can be skipped)	118.1 µs/cast	119.2 µs/cast

Same results. I hashed every result (time of impact, witnesses, normals, status) of 200k random casts in 2D and in 3D, over random heightfields from flat to rough, before and after: identical. The pre-check is bounded by max_time_of_impact only, not by the best hit so far. With the tighter bound, 16 of the 200k casts on flat heightfields picked a different triangle among near-ties, because GJK's time of impact can be a hair earlier than the exact Aabb one.

The new tests compare single-cell (2D: single-segment) heightfields against shape-casting their triangles one by one, and fail if the pre-check is made 10% too tight. They use single cells because larger heightfields don't match brute force even before this change: the traversal sometimes misses triangles. I'll open a separate issue for that.