Simplify grasp verification

dimos/manipulation/grasp_verification.py does more than the problem needs, and the extra machinery is where its bugs come from.

The module answers one question: after commanding a close, is something in the jaws? The physics is a single comparison on the settled readback — above closed_position + empty_epsilon means the jaws stalled on something. Everything else exists to work around the fact that GripperControlTask latches a target and measures position every tick but never reports "I am done moving", so await_gripper_settle reconstructs that by RPC-polling position and guessing with a moved or arrived condition.

That guess is the source of both bugs we have hit. Judging arrival by settle_tolerance means a gripper resting against its mechanical stop (the xArm sits at 0.988 of a nominal 850 count) can never "arrive", so re-opening an already-open gripper times out — that is what #3870 patches. Judging it by moved alone has the mirror failure: commanding open on an already-open gripper moves nothing, so it never settles either. Both are latency problems being solved with a position heuristic.

Two other things worth deleting while we are in here:

• Ten config fields, and no blueprint in the repo sets any of them. The only non-definition use anywhere is the open_tolerance line #3870 adds.
• RobotModelConfig.grasp_verification is a second complete copy of that config that nothing ever reads.
• open_failure / open_tolerance is ambiguous by construction — the open and held bands overlap around 0.85-0.95, so readback cannot distinguish "open" from "holding something wide".
• held_high / open_margin ("jaws never closed") largely duplicates the set_gripper_position return check _command_and_settle already does.

Proposed shape: sleep once past the actuation latency, poll for N stable readings, then apply the one threshold. That removes moved, arrived, arrival_tolerance, held_high/open_margin, open_failure and the dead config field, and takes the file from ~180 lines and 10 knobs to roughly 90 and 4.

The better fix underneath is to have GripperControlTask report motion-complete, since it already owns both the setpoint and the measurement. Then pick/place waits on a signal instead of polling position across a process boundary and the loop goes away entirely. That touches a control-loop task used beyond manipulation, so it is worth doing separately.

Blast radius for the manipulation-only version is small: pick_and_place_module.py is the single real consumer, plus its tests and the dead RobotModelConfig field.

Filed while landing #3870, which patches the arrival case rather than removing the guess. #3870 stays as the near-term fix so the xArm grasping stack can complete a pick; this issue tracks doing it properly.