Speed up mju_boxQP ~2x and improve single-precision accuracy.

`mju_boxQPoption` previously refactored the free-subspace Hessian from scratch whenever the active set changed and performed multiple $O(n^2)$ matrix-vector products per iteration and per line-search step. Updating the Cholesky factor incrementally for small active-set changes and evaluating the search direction and Armijo line search in the `nfree` subspace speeds up `mju_boxQP` by 1.4–1.6x on cold starts and 1.9–2.9x on warm starts, while improving single-precision accuracy.

- Add `cholDelete` and `cholInsert` in `engine_util_solve.c` to update `R` via `mju_cholUpdate` when the active-set change is small relative to `nfree`, and fuse the inner column loops in `mju_cholUpdate`.
- Solve `H_free,free * temp = grad_free` directly after checking `norm2 < mingrad` and evaluate the projected Armijo line search in the `nfree` subspace using `R`, removing `mulVecMatVecSym`.
- Enable `BoxQPTest.BoundedQPvariations` under single precision in `engine_util_solve_test.cc` and evaluate coordinate nudges in `isQPminimum` directly from the gradient and diagonal curvature.
- Rename `lu_benchmark_test` to `engine_util_solve_benchmark_test` and add `BM_BoxQP` benchmarks for $n \in \{10, 30, 100, 300\}$.

PiperOrigin-RevId: 1004087455
Change-Id: Ib416aa55945f543652e2a73f1d217e6761036613

Author: yuvaltassa

doc/changelog.rst

@@ -63,6 +63,8 @@ Engine
     matrix in ``qpos0``, and the kinematics in ``qpos_spring``. Each entry is ``mjSC_IN`` if the group reads the array,
     ``mjSC_OUT`` if it writes it, ``mjSC_INOUT`` if both, or 0. ``MJSETCONST_SCALARS`` lists the scalars it
     writes.
+- :ref:`mju_boxQP` now updates its Cholesky factor incrementally with :ref:`mju_cholUpdate`, making cold-started solves
+  1.4-1.6x faster and warm-started solves 1.9-2.9x faster.
 
 Compiler
 ^^^^^^^^