chore(deps): bump unilab-rl to 1.4.1

[TATP-233]

Summary

Bump the optional unilab-rl (uni_rl) runtime dependency from 1.4.0 to 1.4.1 (latest release on PyPI). No driving issue — routine dependency refresh.

unilab-rl 1.4.1 (release notes) is not a pure packaging bump: it replaces the timing/* / lowercase perf/* TensorBoard/wandb tags with a canonical metric schema (retired keys fail closed), removes the reward/reward_metrics logger interfaces, and adds log_interval throttling (UniLab configs already expose training.log_interval). This PR therefore also migrates UniLab's off-policy metric consumers:

• pyproject.toml / pyproject.rocm.toml: uni_rl extra pin and dev-group pin (both must agree or uv lock has no solution)
• uv.lock / uv.rocm.lock: surgical pin update to unilab-rl 1.4.1 (version, specifiers, sdist/wheel URLs+hashes from a real uv lock resolve; no resolver churn)
• tests/utils/test_experiment_tracking.py, tests/algos/test_offpolicy_dp_sync.py: 12 tests updated to the new logger API — log_collector() without mean_reward, return_mean_ep100 instead of reward=, canonical Perf/* / Loss/* / Train/* keys, seconds-vs-milliseconds unit changes, retired-key negative assertions
• scripts/benchmark/rl/extract_offpolicy_metrics.py, benchmark_offpolicy_dp_scaling.py: tag maps migrated with newest-first (tag, scale) candidates so historical (pre-1.4.1) event files still extract; retired perf/effective_samples_per_sec is derived from run config + Perf/iteration_time when absent (learner_throughput_source records the origin); two retired-without-replacement rows report NaN on new runs by design
• docs/sphinx/source/{en,zh_CN}/2-user_guide/1-training/3-logging.md, 2-algorithms/2-appo.md, zh_CN 3-sac.md: bilingual docs migrated to the canonical schema with a 1.4.1 old-to-new tag table and unit exceptions; historical run evidence in 5-genesis.md / 7-newton.md / 5-support_matrix.md intentionally untouched

src/ needed no changes: UniLab runtime code does not call the logger APIs directly (verified by grep).

Linked Work

• Issue: none — routine dependency refresh
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created or updated — see note below on the 2 pre-existing asset failures
• Additional task-specific validation listed below

Commands actually run (final local head c3c0465):

uv lock --check                      # main lock; ROCm lock checked via pyproject swap
uv sync --extra mujoco --extra motrix --extra uni_rl
uv pip show unilab-rl                # 1.4.1
make check                           # PASS: ruff, mypy 134 files, pyright 0 errors
uv run pytest -m "not slow" -q       # 1557 passed, 27 skipped; 2 pre-existing failures
uv run pytest tests/benchmark/ -q    # 78 passed, 1 skipped
uv run pytest tests/scripts/test_check_docs.py -q   # 19 passed
UNILAB_DOCS_SKIP_AUTODOC=1 sphinx-build -b html -n  # no new warnings

The 2 remaining failures (tests/base/test_sim_backend_smoke.py: allegro_hand YCB meshes not downloaded from the asset hub locally) are the same pre-existing failures documented on main in #1640/#1643, unrelated to this change. Full make test-all locally is blocked by the same missing assets; relying on current-head remote CI as the gate per the contributing workflow.

Remote CI route:

• Base main: current-head CI recorded below once the run starts.

Impact

• Backend impact: none (packaging + logging-consumers only)
• Platform impact: Linux (uni_rl wheels are Linux x86_64); macOS unaffected
• Training effect expected: yes — training runs on 1.4.1 emit the canonical TensorBoard/wandb metric schema; dashboards/queries keyed on retired timing/* / perf/* tags must use the new tags (migration table in docs/sphinx/source/en/2-user_guide/1-training/3-logging.md). FastSAC/FlashSAC/WarpSAC on NVIDIA CUDA move to a single whole-update-cycle CUDA Graph per the upstream release notes.

Artifacts

• N/A

Checklist

• Added or updated tests where needed (12 tests migrated; 7 new benchmark-extraction tests)
• Updated docs if behavior or workflow changed (bilingual logging/algorithm pages)
• Linked the driving issue — none; routine refresh
• Noted any follow-up work explicitly — derived learner throughput in benchmark_offpolicy_dp_scaling.py is an approximation (rank-0 Perf/iteration_time only), documented in the script docstring