feat: integrate IsaacGym fixed model variants (#1577)

[TATP-233]

Summary

• Consume the published IsaacGym fixed-variant adapter by raising the base UniSim requirement to unisim-core>=1.4.1 in the default and ROCm dependency profiles. The optional superdex extra keeps its own >=1.4.0 constraint: SuperDex has no 1.4.1-specific change, and the base requirement already forces every installed profile to 1.4.1 or newer. UniLab's package version is unchanged.
• Extend the deterministic IsaacGym protocol mock to validate variant sources, assignment bounds, per-variant dof-field tables, assigned keyframes, and handshake assignment echo.
• Add IsaacGym fixed-variant coverage:
  • default test lane through the real wire protocol and shared memory;
  • public-layout drift fail-closed test;
  • real Preview-4 slow lane for actor-level asset selection, per-variant defaults, finite rollout, and per-env playback source.
• Add the IsaacGym fixed-variant scale benchmark. Each variant count runs in a fresh child process and reports source generation, construction time, live worker RSS, and control/physics/env-step rates.
• Document the IsaacGym fixed-variant contract and explicitly keep production 600-tool support gated on the real tool-catalog benchmark.

Closes #1578.
Roadmap: #1577.
Benchmark decision: #1579.

Upstream release evidence

• feat(isaacgym): implement fixed model variants (#77) unilabsim/unisim#78 merged at d4ee22f717ba0160178a2d7e7bf786702dd7420a.
• Merged-main ci.yml run 34946318565: success.
• Tag v1.4.1 release run 34946437978: build, sdist verification, and PyPI trusted publishing succeeded.
• PyPI metadata and sdist unisim_core-1.4.1.tar.gz verified.

Validation

Commands run on final local head after installing the published dependency:

uv sync --locked --extra mujoco --extra motrix
uv lock --check
make test-all
uv run --no-sync pytest -m slow tests/base/test_isaacgym_fixed_variants.py -q
uv run --no-sync pytest tests/scripts/test_check_docs.py -q

Results:

• make test-all: PASS
  • formatting/lint/type gates passed (mypy 0 issues; pyright 0 errors, 1 pre-existing optional Drake import warning);
  • non-slow coverage suite: 1521 passed, 27 skipped;
  • benchmark smoke: module mode 34/35 and script mode 35/36 (one platform-optional MLX skip each).
• Real IsaacGym Preview 4 slow lane: 1 passed on RTX 4090 with published unisim-core==1.4.1.
• Docs gate: 19 passed.
• uv lock --check: PASS.

Benchmark results and boundaries are recorded in #1579 (comment).

Impact

• Backend impact: IsaacGym integration/tests/benchmark only.
• Platform impact: Linux IsaacGym runtime lane; protocol tests run without the SDK.
• Training effect expected: no production task owner changes in this PR.
• Support claim: none added for the real 600-tool SimToolReal catalog; that remains gated on Benchmark IsaacGym fixed-variant scale and decide support #1579.

Artifacts

• W&B: n/a
• benchmark result: #1579 (comment)
• video / screenshot: n/a
• ONNX / checkpoint: n/a

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue (Cover IsaacGym fixed variants through protocol mock and gated runtime #1578; roadmap roadmap：IsaacGym 后端 per-env fixed model variants（actor-level asset selection） #1577)
• Noted follow-up work explicitly: production tool-catalog benchmark/support decision stays in Benchmark IsaacGym fixed-variant scale and decide support #1579