replace SPARSE_CONSTRAINT_JACOBIAN with m.is_sparse

[thowell]

almost ready to enable the sparse constraint jacobian with m.is_sparse

(don't think these todos need to be block for this pr)
todos:

• feature to specify number of non-zeros in constraint jacobian set constraint jacobian nonzeros njmax_nnz #1221
• heuristic for a default value for number of non-zeros less than njmax * nv heuristic for estimating the number of non-zeros in constraint_jacobian #1241
• improve elliptic friction cone performance update update_gradient_JTCJ_sparse #1225

---

three humanoids

mjwarp-testspeed benchmarks/humanoid/three_humanoids.xml --nworld=8192 --nconmax=100 --njmax=192

main eaa287f

• SPARSE_CONSTRAINT_JACOBAIN=False so the dense efc_J code path is utilized

Loading model from: benchmarks/humanoid/three_humanoids.xml...

Model
  nq: 84 nv: 81 nu: 63 nbody: 49 ngeom: 58
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: True
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 819200 njmax: 192
Rolling out 1000 steps at dt = 0.005...

Summary for 8192 parallel rollouts

Total JIT time: 0.43 s
Total simulation time: 16.19 s
Total steps per second: 506,108
Total realtime factor: 2,530.54 x
Total time per step: 1975.86 ns
Total converged worlds: 8192 / 8192

this pr

• sparse constraint jacobian is enabled by default via m.is_sparse

Loading model from: benchmarks/humanoid/three_humanoids.xml...

Model
  nq: 84 nv: 81 nu: 63 nbody: 49 ngeom: 58
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: True
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 819200 njmax: 192
Rolling out 1000 steps at dt = 0.005...

Summary for 8192 parallel rollouts

Total JIT time: 0.45 s
Total simulation time: 13.07 s
Total steps per second: 626,670
Total realtime factor: 3,133.35 x
Total time per step: 1595.74 ns
Total converged worlds: 8192 / 8192

[Kenny-Vilella]

I investigated from newton side the effect of the sparse Jacobian:
Used 6 cases spanning DoF Numbers from 36 to 135, for most cases, several units of the same robot were added to get higher DoF count.

Results is that the new sparse version is 20-40% faster for both Newton and CG solver and for both pyramidal and elliptic cone friction.
All the UTs are also passing.
Great work !

I am quite sure that it also gives us a lot of new opportunity for optimization.

[thowell]

minor bug fix, needed to zero the solver context newton solver hessian ctx.h.zero_(). pr description is updated with updated results.

[erikfrey]

Heck yes - ship it!