Add energy rewards and MujocoStyleActuatorCfg (closes S4)

[KraHsu]

Summary

Phase S4 of the Genesis 1.0 migration (ADR-0001). Ships the two S4
sub-tasks that have a clean v1.0 implementation: the energy rewards and
a Mujoco-style actuator cfg helper. The remaining two boxes
(SimulationCfg.actuator_model selector, zero-copy writer refactor)
are explicitly closed as not applicable on v1.0 — see the drift
table below.

Changes

Energy rewards

• src/genelab/mdp/rewards/energy.py — three new reward primitives
  wrapping Genesis 1.0's per-entity energy accessors:
  • kinetic_energy_l2(env, asset_cfg=None) — T² per env, non-negative.
  • potential_energy(env, asset_cfg=None) — signed potential energy.
  • energy_budget(env, target_total=0.0, asset_cfg=None) — squared
    deviation of total mechanical energy from a target.
• tests/test_mdp_rewards_energy.py — synthetic _FakeGsHandle
  exercises every reward path (4 tests).

Mujoco-style actuator cfg

• src/genelab/actuator/mujoco_style.py — MujocoStyleActuatorCfg
  accepts Mujoco's (gear, bias_prm, dyntype="none", gaintype="fixed", biastype="affine") shape and translates it into
  (stiffness=-gear*bias_prm[1], damping=-gear*bias_prm[2]) on top of
  IdealPDActuator. Every other dyntype / gaintype / biastype
  combination raises with the same wording Genesis's own MJCF parser
  uses; non-zero bias_prm[0] also raises because IdealPDActuator
  has no constant-offset force term.
• Registered in src/genelab/actuator/__init__.py.
• tests/test_mujoco_style_actuator.py — 8 cfg-translation tests
  (sign convention, every rejection path, pass-through fields).
• docs/concepts/actuators.{en,zh}.md — new "MJCF-style actuator
  configuration" section + table-row pointer at the top.

Shared

• CHANGELOG.md — ### Added entries for both the energy rewards
  and the actuator cfg.

v1.0 drift on the closed S4 sub-tasks

Probing real Genesis 1.0 surfaced that two of the four S4 ADR
sub-tasks don't apply as the ADR scoped them, even after this PR:

ADR sub-task	v1.0 reality	This PR closes it as
New mujoco_general.py runtime actuator class	No runtime "general actuator" class in v1.0. dyntype/gaintype/biastype only appear in genesis/utils/mjcf.py as XML-parse-time concepts; the parser already translates them into the existing setters at load time. There's no separate Genesis class to wrap.	Reinterpreted as MujocoStyleActuatorCfg — a thin Python-side cfg helper that mirrors the same translation MJCF does, so robots configured in Python rather than from an MJCF can use the same named knobs.
SimulationCfg.actuator_model selector	Becomes meaningful only if a new actuator runtime type lands. Without one, the existing ArticulationCfg.actuators field already routes per-joint actuator selection.	Closed as n/a. MujocoStyleActuatorCfg slots into the same ArticulationCfg.actuators dict as every other actuator cfg; no new selector needed.
Zero-copy set_pos / set_quat in _articulation_writer.py	The writer already calls Genesis setters directly with no copies on the hot path (S1 dropped the historical try/except envs_idx fallback). Remaining .contiguous() / .clone() in reset() exist for semantic reasons (Genesis expects contiguous; clone breaks aliasing with _joint_pos_target).	Closed as "no work to do on v1.0". Recommend re-opening only if a profiler under high reset frequency surfaces these as a hot path.
Comparison reference run dc_motor vs mujoco_general	Hinged on the runtime general-actuator class that v1.0 doesn't have. With the cfg-helper interpretation, MujocoStyleActuatorCfg produces an IdealPDActuator — the same runtime the comparison would have been against.	Closed as redundant. A dc_motor vs MujocoStyleActuatorCfg → IdealPDActuator comparison is dc_motor vs IdealPDActuator, which is well-trodden ground in the locomotion best-practices guide.

Drift findings recorded in saved memory
genesis-1.0-migration/s4-api-drift for future phases / the
promotion-PR discussion.

Test plan

• ruff check
• ruff format --check
• pyright — 0 errors, 0 warnings
• pytest — 598 passed (+12: 4 energy + 8 mujoco-style), 0 failed
• mkdocs build --strict

Risks

• MujocoStyleActuatorCfg is stricter than the underlying MJCF
  shape. A non-zero bias_prm[0] raises rather than approximating
  the constant offset as a static torque. Genesis's parser logs a
  warning for the same shape; this cfg promotes that warning to a
  hard error so the user knows the cfg isn't faithful to the MJCF.
• Energy semantics depend on the world-frame z = 0 convention.
  Rewards that want base-height tracking should still prefer
  base_height_l2. Documented in the per-function docstring.

Notes

• ADR-0001 Phase S4 sub-task table updated: actuator + energy boxes
  ticked; actuator_model selector and zero-copy refactor closed as
  n/a-on-v1.0 with the rationale in the row footers. Comparison
  reference-run box closed as redundant given the cfg-helper
  reinterpretation. After this merges, S5 is the only remaining
  phase before the dev/bump-genesis-1.0 → main promotion PR.