Avoid redundant ThreadMap construction for multi-node continuity penalties

[mickaelbegon]

Problem

For multi-threaded direct-collocation OCPs, continuity penalties may build repeated ThreadMap wrappers during penalty preparation even though threaded dispatch consumes the representative mapped function only.

Proposed change

For multi-threaded STATE_CONTINUITY penalties whose node range begins at zero, construct the mapped function and mapped weighted function only for the representative node. Keep the regular per-node non-threaded functions and preserve current behavior for other penalty types.

Acceptance criteria

• unchanged objective, constraints, Jacobian and Hessian on a collocation OCP;
• unchanged IPOPT trajectory;
• regression test that redundant wrappers are not created;
• no persistent cache, avoiding graph invalidation risks.

A downstream minimal fix and tests are being prepared for a future pull request.

[mickaelbegon]

A minimal downstream implementation is available for review on the fork branch https://github.com/mickaelbegon/BiorbdOptim/tree/codex/threadmap-continuity-cache (commit 3342409). It limits the change to multi-threaded non-multinode STATE_CONTINUITY penalties starting at node zero, retains non-threaded local functions, and avoids persistent graph caching. The branch includes regression coverage for collocation and RK4, function/derivative equivalence, sparsity and map reconstruction when the thread count changes.

[mickaelbegon]

Downstream FHO100 benchmark: compiled/cached local STATE_CONTINUITY kernels preserve exact derivatives and global sparsity (1,629,207 Hessian nonzeros; maximum error 2.33e-10). Warm build improves from 295.5 s to 196.1 s; constraints from 35.5 ms to 25.4 ms and Jacobian from 1.30 s to 1.08 s. The full Hessian improvement is only 1.5% (2.714 s to 2.673 s), so this is not yet a reason to replace the native global MX Hessian. The benefit is in local-kernel reuse and first derivatives; global MX AD/assembly remains the limiting path.