Report other memory, e.g. inline allocation.

[erikfrey]

This solves the problem addressed by #925 but in a different way.

Example output of mjwarp-testspeed benchmark/aloha_pot/scene.xml --nconmax=24 --njmax=128 --memory:

Loading model from: benchmark/aloha_pot/scene.xml...
  nbody: 26 nv: 23 ngeom: 204 nu: 14 is_sparse: False graph_conditional: True
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
  solver: NEWTON iterations: 100 linesearch: iterative ls_iterations: 50
  cone: ELLIPTIC integrator: EULER
Data
  nworld: 8192 naconmax: 196608 njmax: 128

Rolling out 1000 steps at dt = 0.002...

Summary for 8192 parallel rollouts

Total JIT time: 0.71 s
Total simulation time: 4.09 s
Total steps per second: 2,002,968
Total realtime factor: 4,005.94 x
Total time per step: 499.26 ns
Total converged worlds: 8192 / 8192

Model memory 5.38 MB (0.25% of used memory):
 (no field >= 1% of used memory)
Data memory 427.43 MB (19.55% of used memory):
 geom_xmat: 57.38 MB (2.62%)
 efc.J: 96.00 MB (4.39%)
Other memory: 1753.25 MB (80.20% of used memory)
Total memory: 2186.06 MB (4.50% of total device memory)