Solver performance optimization: selective joint computation

[yutingye]

Summary:
Re-applies the solver optimization idea from D51049715 on the current master.

Currently, SkeletonState computes both transforms and derivatives for all joints during solver evaluations. This is wasteful when only a subset of joints is relevant to the error functions being solved.

This diff adds an optimization to:

1. Introduce a JointSet type (bitset<1024>) for efficiently tracking sets of joints.

2. Add getAffectedJoints() virtual method to SkeletonErrorFunctionT that returns which joints are directly used by each error function. Error functions can override this to provide an accurate list (default returns all joints set).

3. Add an initialize() method to SolverFunctionT that is called once before each solve loop. In SkeletonSolverFunctionT, this pre-computes:

   • activeJointXform_: joints that need transforms (all ancestors of affected joints)
   • activeJointDeriv_: joints that need derivatives (only active ancestors of affected joints)

4. Add a SkeletonStateT::set() overload that accepts per-joint needXform and needDeriv flags, allowing selective computation.

5. Add parameter caching in updateSkeletonState() to skip redundant state updates.

6. Implement getAffectedJoints() for CollisionErrorFunctionT and MarkerErrorFunctionT to provide accurate affected joint lists.

As a side effect, setEnabledParameters() now properly propagates to all error functions at the beginning of each solve via initialize(), fixing a long-standing issue where downstream callers had to ensure setEnabledParameters() was called after all error functions were added.

Differential Revision: D95663660

[meta-codesync]

@yutingye has exported this pull request. If you are a Meta employee, you can view the originating Diff in D95663660.