feat: consume unisim 1.7.5 physics-state playback contract in interactive rendering

[TATP-233]

Summary

• Downstream adaptation for feat: 统一物理状态回放渲染契约，支撑 viser 等渲染前端适配全部回放后端 unilabsim/unisim#291 on the released unisim-core 1.7.5 playback contract: play_interactive / play_viser now precheck supports_physics_state_playback at startup and fail with an actionable message (naming the backend and its available rendering paths) instead of surfacing NotImplementedError from the first snapshot fetch inside the viewer loop.
• New owner module unilab.visualization.playback_state: PhysicsStateApplier splits snapshots through the contract physics_state_layout().split_state() — the hardcoded mjSTATE_FULLPHYSICS / 1+nq+nv slicing is gone from both entrypoints — and replays mocap geometry through the contract get_playback_mocap_state entrypoint (aligned with the playback model; falls back to the snapshot tail only when the backend declares no mocap capability but the layout carries the columns).
• EnvPlayCapabilities gains supports_mocap_playback; NpEnv exposes get_physics_state_layout() / get_playback_mocap_state() passthroughs with fail-closed gating.
• Dependencies: unisim-core>=1.7.4 → >=1.7.5 (playback contract release) and mjbatch-uni~=0.2.3 → ~=0.2.4 (streaming VariantPack.builder, required by unisim-core 1.7.5's fixed-variant check — without it tests/base/test_entity_scene_consumer.py fails closed).

User-facing impact: backends without the playback contract (genesis/isaacgym) now get a clear startup error pointing at alternatives; mocap-body tasks render correctly in interactive viewers instead of crashing on mj_setState size mismatch. No training-path behavior change.

Linked Work

• Issue: feat: 统一物理状态回放渲染契约，支撑 viser 等渲染前端适配全部回放后端 unilabsim/unisim#291 (downstream companion of unisim PR fix: remove unnecessary np.asarray in motrix backend & bump motrixsim #293, released in unisim-core 1.7.5)
• Parent roadmap (when applicable): —
• Roadmap declared base (when applicable): —
• Milestone: —
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created or updated
• Additional task-specific validation listed below

Commands actually run:

uv run --no-sync ruff format                       # 381 files unchanged
uv run --no-sync ruff check --fix                  # all checks passed
uv run --no-sync mypy src/unilab                   # 132 source files, no issues
uv run --no-sync pyright                           # 0 errors (1 pre-existing drake_uni import warning in cli.py)
uv run --no-sync ruff check tests --select F401,F821,F811,F841 --output-format concise
uv run --no-sync pytest -m "not slow" --cov=src/unilab --cov-report=term-missing -q
#   1539 passed, 26 skipped; the single failure was
#   tests/scripts/test_check_docs.py flagging *locally generated, gitignored*
#   scripts/benchmark/outputs/ artifacts (stale `_num_envs` key from a same-day
#   local benchmark run) — the test passes on a clean tree (19/19) on both main
#   and this head; not caused by this PR.
uv run --no-sync python scripts/benchmark/smoke_test.py   # 35/35 modules, 36/36 entrypoints

Focused suites on the final head (unisim-core 1.7.5 + mjbatch-uni 0.2.4):

uv run --no-sync pytest tests/visualization/ tests/base/ -q
#   582 passed, 17 skipped; 2 failures are a pre-existing test-isolation bug
#   (tests/visualization/test_playback_session.py pollutes
#   tests/base/test_mjwarp_playback.py when run in the same session) —
#   reproduced with this PR's changes stashed; out of scope here.

Remote CI route:

• Base main: current-head (fb0347e0) CI all green — https://github.com/Motphys/UniLab/actions/runs/35863164309 (ruff-format 3m15s, ruff-lint 1m58s, mypy 40s, pyright 1m25s, test (ubuntu-slim) 5m29s, benchmark-smoke 3m44s) and https://github.com/Motphys/UniLab/actions/runs/35863164396 (Build Sphinx 38s).

Impact

• Backend impact: mujoco (MuJoCo playback-shell rendering path consumed by play_interactive / play_viser; mjwarp/newton/drake/superdex snapshots flow through the same contract entrypoints)
• Platform impact: Linux (verified); macOS unaffected paths
• Training effect expected: no

Artifacts

• W&B: —
• benchmark result: —
• video / screenshot: —
• ONNX / checkpoint: —

Checklist

• Added or updated tests where needed (tests/visualization/test_playback_state.py, tests/base/test_np_env_playback_contract.py)
• Updated docs if behavior or workflow changed (no user workflow change; entrypoint behavior is fail-fast with clearer errors)
• Linked the driving issue
• Noted any follow-up work explicitly: genesis/isaacgym playback adaptation is tracked separately per feat: 统一物理状态回放渲染契约，支撑 viser 等渲染前端适配全部回放后端 unilabsim/unisim#291 item 4; the test_playback_session.py → test_mjwarp_playback.py test-isolation bug and the stale-benchmark-output docs-check sensitivity are pre-existing issues observed during validation and left untouched.