fix: align genesis probe parallel level

[TATP-233]

Summary

• Temporarily set GS_PARA_LEVEL=2 around SAC/TD3/FlashSAC runner construction when training.sim_backend is genesis.
• The one-env learner probe now compiles the same Genesis/Quadrants kernel lane as the scaled collector; the prior value or unset state is restored before runner.learn().
• Other backends remain untouched. This is a cold-start mitigation and does not change the collector-ready root fix or claim new convergence evidence.

Linked Work

• Issue: roadmap: g1_motion_tracking multi-backend enablement — newton CUDA graph, genesis cold-start, isaac worker fixes #1604 (Genesis cold-start mitigation work item)
• Parent roadmap: roadmap: g1_motion_tracking multi-backend enablement — newton CUDA graph, genesis cold-start, isaac worker fixes #1604
• Roadmap declared base: main
• Milestone: none
• 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 pytest tests/scripts/test_train_scripts.py -k "genesis_probe_para_level"
# 147 deselected because this file is in the slow lane; rerun below with -m slow
uv run pytest tests/scripts/test_train_scripts.py -m slow -k "genesis_probe_para_level"
# 3 passed, 144 deselected
uv run ruff format --check src/unilab/scripts/train_offpolicy.py tests/scripts/test_train_scripts.py
uv run ruff check src/unilab/scripts/train_offpolicy.py tests/scripts/test_train_scripts.py
make check
make test
# Initial run failed during collection before optional extras were installed.
make sync
make test
# Initial clean-worktree run had two missing Allegro asset-cache failures.
uv run unilab-pull-assets --robot allegro_hand
uv run pytest tests/base/test_sim_backend_smoke.py::test_mujoco_backend_fixed_base_dof_views_do_not_skip_first_joint tests/base/test_sim_backend_smoke.py::test_motrix_backend_fixed_base_base_views_are_available -m "not slow"
# 2 passed
make test
# 1485 passed, 28 skipped, 601 deselected
make test-all
# 1485 passed, 28 skipped, 601 deselected; benchmark smoke module-mode 34/34 and script-mode 35/35
uv run pytest tests/scripts/test_check_docs.py -q
# 19 passed
cd docs/sphinx && UNILAB_DOCS_SKIP_AUTODOC=1 uv run --no-project --with-requirements requirements.txt sphinx-build -b html -n source build/html
# build succeeded
git diff --check

Remote CI route:

• Base main: current-head CI passed at commit 2444ec45: https://github.com/Motphys/UniLab/actions/runs/35207321261
• Docs build passed: https://github.com/Motphys/UniLab/actions/runs/35207321297

Impact

• Backend impact: genesis
• Platform impact: Linux
• Training effect expected: yes, reduced first-run Genesis cold JIT/cache mismatch; no reward-parity claim

Artifacts

• W&B: none
• benchmark result: none
• video / screenshot: none
• ONNX / checkpoint: none

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue
• Noted any follow-up work explicitly