Enable viser rendering for the motrix backend

[TATP-233]

Summary

Follow-up to #1638: unisim-core 1.7.6 (unilabsim/unisim#302) implements the physics-state playback contract on the Motrix adapter — for both portable entity scenes and whole-MJCF scenes guarded by a construction-time MJCF joint-order validation — so the browser-based viser viewer no longer fail-closes on --sim motrix.

• src/unilab/cli.py: motrix leaves VISER_UNSUPPORTED_SIMS (genesis/isaacgym/isaacsim keep their fail-closed rejection); viser is browser-based, so macOS motrix playback no longer requires mxpython for this mode.
• Dependency pins move to unisim-core>=1.7.6 (pyproject.toml, pyproject.rocm.toml, uv.lock): the motrix playback contract only exists since 1.7.6.
• Tests: the two former motrix rejection cases become positive routing cases (eval → play_viser.py, train → training.play_render_mode=viser forwarded to the train script); new genesis eval/train rejection cases keep the fail-closed branch covered.
• Docs: en/zh support matrix playback sections record motrix viser routing.

Validation

• uv run pytest tests/test_cli.py tests/visualization -q — 119 passed
• make check — green (ruff, mypy, pyright)
• tests/scripts/test_check_docs.py — only the 3 pre-existing scripts/benchmark/outputs failures already present on clean main
• End-to-end on real checkpoint (logs/rsl_rl_ppo/Go2JoystickFlat/2026-09-22_13-24-20_motrix):
  uv run eval --algo ppo --task go2_joystick_flat --sim motrix --render-mode viser viser.max_envs=2 viser.port=18099 training.play_steps=100
  — viser server comes up, 2 env slots render from motrix physics-state snapshots, ran 150s clean (server loop stopped by timeout).

Known scope boundary (upstream): kinematic mocap mirroring in playback is phase 2, so tasks whose playback MJCF carries mocap bodies (e.g. g1_motion_tracking) still fail closed at the PhysicsStateApplier with an actionable error.

[TATP-233]

CI 暴露了两类问题，均已定位：

1. 上游 1.7.6 缺陷（阻塞本 PR）：extract_mjcf_joint_layout 只读取第一个 <worldbody>，include 合并场景丢关节 → allegro in-hand 场景构造失败。已提上游修复 fix(motrix): merge every worldbody in extract_mjcf_joint_layout unilabsim/unisim#306（CI 绿），等 1.7.7 发版后本 PR 升 pin 到 >=1.7.7。
2. UniLab 测试 fixture 不适配新构造校验：已修复并推送（5e31688e）——stewart 场景 12 个被动铰链命名、fake-model 测试补齐/屏蔽新校验。

本地用修复版 unisim 全量复验：1544 passed，唯一失败是 main 上既有的 docs 检查项。

[TATP-233]

更新：unisim-core 1.7.8 已发布（含 unilabsim/unisim#306 的多 worldbody 合并修复），本 PR 三处 pin 已升到 >=1.7.8 并复锁 uv.lock，CI 全绿。

排查记录：升 pin 后 CI 曾持续报 stewart unnamed hinge on leg00，但 merge ref、PyPI sdist、本地 py3.11 全量复现均正常。根因是测试 job 的 actions/cache（key 只随 hub.py 变化）会把整个 src/unilab/assets/robots 旧快照盖回 checkout，git 跟踪的 scene XML 被悄悄回滚到缓存版本；main 锁 1.7.5 无 joint 校验所以从未暴露。已加 git restore 步骤固定跟踪文件到当前 commit（b6af915f），缓存的 HF 网格文件不受影响。