Remove mjData.qLDiagSqrtInv, add corresponding argument to mj_solveM2.

- `qLDiagSqrtInv` is only required for the dual solvers. It is now computed as-needed rather than unconditionally.
- `mj_solveM2` now requires a new input array `sqrtInvD` which contains the square root of the inverse diagonal D (formerly saved in `qLDiagSqrtInv`).

PiperOrigin-RevId: 710805133
Change-Id: I0622d6a8da3882916824e9c10bad9223c122c321

include/mujoco/mjdata.h

@@ -301,7 +301,6 @@ struct mjData_ {
   // computed by mj_fwdPosition/mj_factorM
   mjtNum* qLD;               // L'*D*L factorization of M (sparse)               (nM x 1)
   mjtNum* qLDiagInv;         // 1/diag(D)                                        (nv x 1)
-  mjtNum* qLDiagSqrtInv;     // 1/sqrt(diag(D))                                  (nv x 1)
 
   // computed by mj_collisionTree
   mjtNum*  bvh_aabb_dyn;     // global bounding box (center, size)               (nbvhdynamic x 6)

include/mujoco/mjxmacro.h

@@ -632,7 +632,6 @@
     X   ( mjtNum,    qM,                nM,          1           ) \
     X   ( mjtNum,    qLD,               nM,          1           ) \
     X   ( mjtNum,    qLDiagInv,         nv,          1           ) \
-    X   ( mjtNum,    qLDiagSqrtInv,     nv,          1           ) \
     XMJV( mjtNum,    bvh_aabb_dyn,      nbvhdynamic, 6           ) \
     XMJV( mjtByte,   bvh_active,        nbvh,        1           ) \
     X   ( mjtNum,    flexedge_velocity, nflexedge,   1           ) \

include/mujoco/mujoco.h

@@ -365,7 +365,8 @@ MJAPI void mj_factorM(const mjModel* m, mjData* d);
 MJAPI void mj_solveM(const mjModel* m, mjData* d, mjtNum* x, const mjtNum* y, int n);
 
 // Half of linear solve:  x = sqrt(inv(D))*inv(L')*y
-MJAPI void mj_solveM2(const mjModel* m, mjData* d, mjtNum* x, const mjtNum* y, int n);
+MJAPI void mj_solveM2(const mjModel* m, mjData* d, mjtNum* x, const mjtNum* y,
+                      const mjtNum* sqrtInvD, int n);
 
 // Compute cvel, cdof_dot.
 MJAPI void mj_comVel(const mjModel* m, mjData* d);