feat: improve spirv-compatibility + bump dependencies + a few bugfixes

[sebcrozet]

Breaking changes

• CompositeShapeRef::project_local_point and project_local_point_and_get_feature now return
  Option<...> instead of unwrapping internally, and take an additional max_dist parameter that
  bounds the search distance.

Modified

• Bump glamx to 0.3 (built on glam 0.33), simba to 0.10, rstar to 0.13, hashbrown to 0.17,
  rand to 0.10, and kiss3d (visual examples) to 0.42.
• Many shape, AABB, and voxel internals were reworked to avoid bracket-indexing of glam vectors,
  which is not supported on SPIR-V targets. A new VectorExt::vget/vset API and a for_each_dim!
  macro are now used in place of v[i]. parry3d's alloc feature also now gates smallvec,
  downcast-rs, rstar, and glamx/approx, so the no-alloc build can target SPIR-V.

Fixed

• Fix multiple EPA failures on large coordinate magnitudes by scaling the face-rejection tolerance
  relative to the simplex vertex magnitudes (#415).
• Fix CCD edge cases (in both ball-vs-ball and support-map-vs-support-map casts) where casts
  starting at (or very close to) the contact boundary with a near-tangent direction would return
  an unreliable normal. The fallback now uses the contact query to recover a robust closest-point
  normal, and the small-TOI threshold was relaxed from 1e-5 to 1e-4.
• Fix Aabb::cast_local_ray_and_get_normal panicking with a subtraction overflow when casting a
  zero-direction ray starting inside the AABB (#383).
• Fix infinite loop in TriMesh::intersection_with_local_plane on degenerate adjacency graphs
  that don't loop cleanly back to the starting index
  (#398).
• Fix panic in mesh_intersection when constraint edges overlap (e.g. for co-planar triangles).
  Overlapping constraints are now skipped instead of crashing the CDT
  (#389).
• Fix panic in 3D voxelization's internal convex-hull step by falling back to try_convex_hull
  and returning an empty hull on failure (#347).
• Fix WSign::copy_sign_to on non-CUDA targets by using the native copysign (the bit-twiddling
  workaround is now scoped to nvptx64, where cuda_std's copysign does not compile).

fix #415
fix #383
fix #398
fix #389
fix #347