fix(isaac): resolve default Franka USD for Isaac Sim 6.0 asset move (closes #110)

[yinsong1986]

What changed

NVIDIA relocated the bundled Franka Panda asset in Isaac Sim 6.0, from the
4.x path {assets_root}/Isaac/Robots/Franka/franka.usd to a vendor folder at
{assets_root}/Isaac/Robots/FrankaRobotics/FrankaPanda/franka.usd. Every
example that loads the default Franka USD 404'd on 6.0 because the sub-path was
hardcoded to the stale 4.x layout.

This PR makes the default Franka resolution version-robust in all three
affected helpers:

• examples/isaac_gs/scene.py → _default_franka_usd() (the reproduced failure)
• examples/libero/run_isaac.py → _resolve_robot_asset()
• examples/libero/run_isaac_agent.py → _resolve_robot_asset()

Resolution logic (shared shape in each file):

1. HEAD-probe the 6.0 vendor path first (current target runtime).
2. Fall back to the legacy 4.x path.
3. On a definitive 404 of all HTTP candidates, raise with a clear
   pass --robot-usd/--robot-urdf (resp. robot_usd=...) hint naming the
   tried paths.
4. A non-HTTP root (e.g. an omniverse:// Nucleus path that can't be
   HEAD-probed) is treated as inconclusive and degrades to the 6.0
   candidate rather than failing on a probe it can't perform — add_robot
   then surfaces a clear error if it's wrong.

Secondary (also from the issue)

render_demo exited 0 even when build_default_scene raised, because
Isaac's SimulationApp fast-shutdown swallows the non-zero exit. The
__main__ entry point now catches, logs, flushes, and forces a non-zero exit
via os._exit(1) after teardown, so CI/scripts notice the failure.

Why

The 4.x → 6.0 asset move is permanent and the bucket layout differs by release,
so a single hardcoded sub-path can't be correct for both. HEAD-probing the
candidates makes the examples run out-of-the-box on whichever Isaac Sim version
the host has, while keeping the explicit --robot-usd / --robot-urdf
overrides as the escape hatch.

Test surface

• hatch run lint — pass (black/isort/flake8 over strands_robots_sim + examples).
• hatch run test — 203 passed, 33 skipped (skips are GPU/Isaac-gated).
• Resolver logic validated against the public Omniverse content bucket:

  200  .../Assets/Isaac/6.0/Isaac/Robots/FrankaRobotics/FrankaPanda/franka.usd   (6.0 — selected)
  404  .../Assets/Isaac/6.0/Isaac/Robots/Franka/franka.usd                        (legacy 4.x)
  None .../omniverse://server/foo.usd                                             (non-HTTP — inconclusive)
  

  _resolve_* correctly selects the 6.0 path.
• AST parse-check on all four edited files.

No new test failures introduced. (The 7 environmental failures in
test_rendering.py / test_policy_runner.py mentioned in the issue context
are not part of this repo's hatch run test slice — that slice is the
strands_robots_sim/isaac/tests/ Isaac unit/lazy-import suite, which is fully
green here.)

Acceptance criteria (from #110)

• Update the default Franka USD to the 6.0 vendor path in all three helpers.
• Make it version-robust: try the 6.0 path, fall back to legacy, HEAD-probe.
• Emit a clear "asset not found — pass --robot-usd/--robot-urdf" hint on a definitive miss.
• Secondary: propagate render_demo failure to the process exit status.

Out-of-scope follow-ups

• None required for this defect. The HEAD-probe adds at most two short HTTP
  round-trips on the default path; if that ever becomes a concern on
  cold-start-sensitive CI, a small per-process cache of the resolved root could
  be a separate optimization.

Closes #110.