Add continuous-collision infrastructure for deformables

[smallquail]

The standard discrete collision pipeline generates contact points at a configuration; this new module prices gaps along trajectories:

• differentiable vertex-triangle / edge-edge / vertex-geom distance kernels with closest-point barycentrics (mjc_PtTri, mjc_SegSeg, mjc_GeomDist, plus sharp-feature extraction mjc_GeomVerts/mjc_GeomEdges);
• swept-volume contact-candidate generation over the flex BVH (mjc_candidates);
• per-pair gap evaluation with the gradient's direction and vertex weights (mjc_pairGap);
• conservative advancement bounding each pair's time of impact (mjc_advance).

Design notes

• mjcPair is the geometric identity of a candidate pair only. Solver state (multipliers, ages) and cached linearizations belong to the consumer, not to this struct — the module neither reads nor writes them.
• No hard-coded constants. The two lengths the module needs are caller-supplied parameters:
  • the standoff cap bounds the rest gap: an active pair settles at a small positive distance δ = min(pair band, cap) rather than at zero. Strictly positive so the conservative advancement always sees a nonzero time of impact (a pair resting at exactly zero would lock the advance); the per-pair min keeps thin participants' skins proportionally thin, and the cap stops thick ones carrying a fat layer.
  • the detection band is the tracking reach, and is not a physical length: how early the broad phase starts holding a pair and how long the gap evaluation keeps it held. Wider costs candidates; narrower hands pairs to the consumer later.
• Flex-flex gaps are measured at the midsurface, not skin-to-skin: where the mesh geometry is tighter than the combined radii (a string threaded through a hem), a skin gap is permanently negative and the pair would be discarded as invalid — losing CCD coverage exactly where tunneling is likeliest. The broad phase adds the radii back into its reach, so detection range is unchanged.
• Engine-internal (MJAPI on entry points only for tests, marked not a supported API). No user-visible behavior change; changelog notes the new infrastructure.

Testing

engine_collision_continuous_test: distance kernels across their regions (interior/edge/vertex, crossing/collinear/parallel), geom signed distance and sharp features, the pair gap with its gradient verified by central differences at every involved vertex, the conservative advancement (analytic cap on a crossing sweep; the conservative closing-rate bound; small-motion early-out), and candidate generation on stacked cloths (pairs found at 2 mm, none at 0.5 m). Full suite: 1349/1349.

[yuvaltassa]

Can you ask @kbayes to review? I feel like the unit tests could be improved and also would his eyes on this re consistency with the rest of the collision code.

[quagla]

Superseded by b924ac6