fix(maniskill): apply ManiSkill gripper grasp friction (2.0) so demo action-replay completes tasks 1:1

Root-cause: ManiSkill sets the panda finger contact material to friction 2.0
(+ patch radius 0.1) in code via Panda.urdf_config, NOT in the URDF, so the
sapien3 loader left the shipped fingers at the scene-default 0.3. The grasped
object slipped ~1 cm during a lift and success flipped near the goal boundary
(PickCube demo-replay 8/12). Replicating the material pulls the contact-phase
object trajectory to the PhysX CPU/CUDA noise floor (~2e-4); PickCube replay is
now 22/25, matching native physx_cpu (24/25) with 23/25 episode agreement.

- recipe.apply_maniskill_gripper_friction(): per-shape PhysxMaterial(2.0,2.0,0)
  + patch radius, idempotent, no-op for gripperless robots (panda_stick).
- base/multi_robot tasks apply it once on reset (opt-in, maniskill-only).
- New render_demo_replay.py: replays an official pd_joint_delta_pos demo through
  BOTH native ManiSkill and the shipped task from the same initial state, and
  only emits a clip where both reach success (render-verify ranked candidates;
  full demo state incl. object velocity to kill cross-rollout leak). Handles
  single-robot, panda_stick (7-dim), and multi-robot (16-dim) layouts.
- Gallery: 11 real demo-replay 1:1 task-completion videos replace the prior
  synthetic-action clips (pick/push/pull/stack/poke/lift_peg/roll_ball/
  pull_cube_tool/stack_pyramid/push_t/two_robot_pick_cube); peg_insertion/
  plug_charger/draw_triangle/place_sphere documented honestly as not shown.
- Regression test test_maniskill_gripper_friction.py (3 tests).

Author: geng-haoran