fix(maniskill): truncate demo-replay clips at task completion — PushT 0.87->0.067/255

The PushT residual was NOT a reproduction gap: at the moment the task succeeds
the Tee divergence is only ~0.3mm — the 6mm came entirely from ~75 post-success
demo steps where the RL policy keeps nudging the already-placed Tee, accumulating
contact drift the viewer never needs to see.

- render_demo_replay: _native_rollout now returns the first success step;
  --truncate-buffer (default 8) ends each clip a few frames after completion, and
  agreement-ranking scores only the shown (pre-completion) window. The clip is now
  the genuine 1:1 completion.
- Deep-verified there is no remaining physical mismatch on PushT: materials (Tee 3.0,
  table 0.3, stick), geometry (cylinder pusher, Tee 2-box + contact/rest offset),
  solver iters 15/1, link masses/cmass/inertia (exact), drive gains, global PhysX
  flags all match; handler correctly skips passive-force under disabled gravity.
  enhanced_determinism is a dead end (physx_cpu already deterministic run-to-run).
- All 11 clips re-rendered, end on completion: every task now 0.012-0.114/255
  (PushT 0.067). 43 maniskill tests pass.

Author: geng-haoran