feat: consume physical entities in manager reset transactions (#1599)

[TATP-233]

Manager-Based object-only reset previously rebuilt a full state row from global defaults, which could reset the robot and broadcast environment zero's variant state. This change binds logical entities to UniSim's public physical layout, stages one selected-entity reset request and restores per-environment state/control defaults through public backend methods.

The A/B audit additionally removes quadratic row lookup and unused whole-batch snapshots from reset staging. Different joint position/velocity subsets merge without losing untouched columns; validation precedes the single native commit. Cold binding rejects logical roots that name descendant bodies, so queries, defaults and root writes refer to the same physical object. Scene/config owners collect all declared physical/variant asset paths. A registered primitive task uses the same pickleable EnvFactory for MuJoCo and IsaacSim, with passive joints adding state without actions.

Fixes #1599. Refs unilabsim/unisim#108/#113 and merged audit unilabsim/unisim#138. Declared/PR base: main; governing ADRs: ADR-0006 and ADR-0007. No training, learner/runner, reward, M3 contact/DR capability or runtime lifecycle change.

Merge status: all final-head checks pass, but main protection requires one approving review and none is present. GitHub rejected ordinary merge; repository auto-merge is disabled. This PR remains open for the required review; branch protections were not bypassed.

Released dependency: normal and ROCm metadata require unisim-core>=1.5.0, and both locks resolve PyPI 1.5.0 without Git source overrides. Upstream v1.5.0 release passed source-distribution installation/smoke and publication. uv sync --locked --extra mujoco --extra motrix installed it in site-packages; distribution metadata is 1.5.0 with no direct_url.json, so this is neither editable nor Git-only validation. ROCm lock resolution is checked, without claiming ROCm hardware execution. IsaacSim recording acceptance was explicitly deferred by the maintainer to unilabsim/unisim#133 and is not claimed passed.

Validation on final implementation/dependency head 9e8a0977:

• make test-all: 1485 passed / 28 skipped / 588 slow deselected, 70% coverage; required benchmark import smoke 34/34 module and 35/35 script, plus Ruff/mypy/Pyright. No UV_NO_SYNC or editable override; optional Drake warning remains.
• UNILAB_TEST_M2_ISAACSIM=1 uv run pytest -q tests/envs/test_multi_entity_consumer.py -k native --basetemp=/tmp/unilab-m2-pypi-native -x: 2 passed with the PyPI dependency, actual registered factory → Manager-Based → public UniSim → native IsaacSim.
• uv lock --check: passed for normal and isolated ROCm profile pairs; git diff --check: passed.
• uv run pytest -q tests/scripts/test_check_docs.py: 19 passed on the final head; CI builds Sphinx.
• Bounded CPU staging A/B has 24 equal-output cases, separating old code, row-index-only optimization and sparse-field optimization. N=R=4096 pose staging: 23.456 → 1.697 → 1.407 ms; no engine/IPC/rollout-throughput claim. Reproduce with uv run python scripts/benchmark/physics/m2_reset_ablation.py --output /tmp/reset-ab.json; upstream bilingual audit records the evidence and limits.
• All six current-head CI jobs passed on 9e8a0977ba73296549addba66dc8f5de161c2752: format, lint, mypy, pyright, benchmark smoke and tests. Sphinx build passed; PR deployment is intentionally skipped. An earlier duplicate run on the same commit was superseded when the PR became ready.

Go2/Allegro assets are obtained via the registered hub and stay ignored. The fresh-process physics/registry test explicitly disables graphics before interpreter startup, avoiding unrelated OSMesa dependencies while still exercising real physics construction/reset. This does not change rendering acceptance.