sim/mujoco: profile and improve run_policy mock-policy throughput (T26)

[cagataycali]

Deferred from PR #85. Finding showed 500 steps in 14.6s for a mock policy at control_frequency=50Hz. 500 steps / 50Hz = 10s real-time target, so ~4.6s of Python overhead per rollout.

DoD

• Profile PolicyRunner.run with a mock policy (e.g. py-spy or cProfile).
• Identify the hotspot: observation build? action mapping? logger/hook dispatch? sim lock contention?
• Goal: 500 steps at 500Hz in <2s with fast_mode=True (batch-eval / data-collection path, no real-time throttling).
• Add a perf regression test under tests_integ/ that pins the wall-clock ceiling.

Notes from PR #85

For now, the run_policy / start_policy docstrings have been updated
(T25) to point callers at fast_mode=True for non-real-time usage. The
performance budget DoD (< 2s) isn't met yet but isn't a correctness bug —
it's a throughput optimisation opportunity.

Tracked per AGENTS.md rule: the project board is the source of truth for
follow-ups.

[cagataycali]

Closed in PR #85 (commit 53fcf47).

Result: 19x speedup, budget met.

cProfile revealed that ~93% of wall time in the T26 target scenario was spent inside mjr_render + mjr_readPixels in _get_sim_observation. The MockPolicy never consumes image obs — the renders were dead work on every step.

Fix:

• New Policy.requires_images: bool property (default True). Policies that only consume joint state (MockPolicy, pure-IK, scripted trajectories) override to False.
• SimEngine.get_observation(skip_images=True) — new keyword-only param; PolicyRunner.run() and .evaluate() propagate the hint from the policy.
• _get_sim_observation(skip_images=True) early-returns after the joint-state dict, before the camera render loop.
• Safety: MuJoCo's get_observation force-overrides skip_images=False when a dataset recorder is attached (recording needs every frame).

Perf (500 steps @ 500Hz, fast_mode=True, so100):

scenario	before	after
fast_mode=True	0.38s	0.02s (19x)
fast_mode=False (real-time)	2.40s	1.50s

Regression tests: TestT26PerfBudget::test_mock_policy_500_steps_under_budget pins the 2s ceiling, and test_requires_images_propagates_to_observation pins the plumbing via a spy on get_observation.