Make the Push EE yaw offset a per-robot property

[amburger66]

Problem

Which face of the gripper leads into the pushed object is a fact about the hand, not the task — but it was a single global flag, skill_push_ee_yaw_offset = 0.0, tuned for the Fetch.

On the Franka that 0.0 turns the Franka Hand so its body reaches the domino before the intended contact face does. It wedges into the block and launches it rather than tipping it — which still satisfies Toppled, so it reads as a success in logs. Caught by rendering the push on a captured domino scene.

Fix

• SingleArmPyBulletRobot.push_ee_yaw_offset — new property, defaults to 0.0 (side push, the Fetch-verified behavior).
• PandaPyBulletRobot overrides it with np.pi / 2 (front push, knuckles spanning the object's face).
• skill_push_ee_yaw_offset becomes None; when set it still forces one offset for every robot, so the two can be compared.
• push.py resolves via a small resolve_ee_yaw_offset(config) helper.

Blast radius

pybullet_robot defaults to "fetch" and only the domino_real env config selects "panda", so coffee / fan / boil are unchanged (0.0 before and after).

Verification

• Replay of a push on the start domino of a captured scene, with no override: reproduces the front push rather than the side push.
• mypy, pylint, yapf, isort, docformatter clean.
• tests/test_skill_factories.py, tests/pybullet_helpers/robots/test_panda.py, tests/envs/test_pybullet_domino_{composed,real}.py, tests/ground_truth_models, tests/test_skill_factories_integration.py — all pass.

[yichao-liang]

Looks good! I used a side push rather than a front push for Fetch because it resulted in fewer collisions. But I guess the shape of the Panda’s gripper makes side pushing difficult?

[amburger66]

I think they're both side pushes... I think the coordinate axes of the Panda hand and the Fetch gripper may be different