Add swept SDF collision sequence error function (anti-tunneling) (#1526)

[cdtwigg]

Summary:

Add SDFCollisionSequenceErrorFunctionT, a two-frame SequenceErrorFunctionT that penalizes a vertex sweeping through an SDF collider between consecutive frames. The single-frame SDFCollisionErrorFunctionT only sees penetration at the sampled poses, so a vertex moving fast enough to pass entirely through a thin collider (a finger, say) in one timestep is never penalized — it tunnels. This closes that gap for the sequence solver, which is used to post-process existing motion.

How it works:

• For each participating vertex, the segment connecting its position at frame t and t+1 is swept against each collider in the collider's local SDF frame: each endpoint is mapped through its own frame's world-to-collider transform and the two local positions are interpolated linearly. This captures the relative motion when both vertex and collider are joint-driven, keeps the interpolation exactly differentiable (no slerp-derivative approximation), and makes the sphere-trace that locates surface crossings exactly conservative (distances and steps are in the same local units, so a thin feature can't be stepped over).
• The penalty is w * phiMax^2 per penetrating sub-interval, where phiMax is the greatest penetration depth reached along the sweep. It is a max, not an integral, so it depends only on how deep the pass-through gets, never on how long the segment stays inside (no degenerate "stretch the trajectory to dilute the penalty" direction). The deepest point is located exactly by bisecting d'(s) = 0; because it is a true interior extremum, the envelope theorem makes the gradient independent of how the deepest point drifts with the parameters, so it reduces to the single-frame contact gradient evaluated there, split across the two frames by the interpolation weights (1 - s*) and s*. One residual per interval; gradient and Jacobian both finite-difference clean.
• Only sub-intervals bounded by two real surface crossings (entry and exit) are penalized, targeting the pure pass-through case and leaving at-frame penetration to the per-frame error function.
• Derivatives reuse the shared detail_sdf_collision chain-rule helpers from the previous diff, evaluated against both frames' skeleton states. The local SDF gradient is mapped to world space including the world-to-collider scale factor, which the derivatives require whenever the collider's joint carries a non-unit scale.

The penetration normal is treated as frozen per evaluation (the standard Gauss-Newton contact linearization), so no SDF second derivatives are needed — only transform derivatives, which the shared helpers already provide.

Known limitation / follow-up: a vertex passing exactly through a collider's medial axis (dead-center through the axis of a thin tube, or perpendicular through a thin sheet) lands on a degenerate point where the SDF normal is undefined and the depth has a non-smooth maximum, so the gradient there is ~0 and gives no signal. This is a measure-zero case for convex colliders (any off-axis pass resolves correctly by deflection); a retract-along-the-entry-normal fallback for it is a planned follow-up. There is also no broadphase culling yet (every participating vertex is swept against every collider) and no pymomentum binding (the single-frame function has one).

Reviewed By: yutingye

Differential Revision: D108807348

[meta-codesync]

@cdtwigg has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108807348.