mcp: measure the TSF-85 tactile constants on real pads

[jlongvalRobotiq]

The tactile datasheet (mcp/gripper_mcp/datasheets/tactile/robotiq_tsf_85.yaml) is tuned against the Isaac twin. Eric's bench numbers on a real 2F-85 + TSF-85 (2026-09-10): rest counts ~12540 per taxel, rest noise peaks ~0.035 of full scale, pads touching each other 0.236, internal hold 0.104, external hold 0.817. #65 works around the gap by deriving the runtime threshold from the tare (max(contact_threshold, noise_margin x measured rest noise)), but the constants themselves are still unmeasured.

When TSF pads are available here:

• Run gripper_tare_tactile and gripper_read_tactile on an empty open gripper, a closed empty gripper, an external hold and an internal hold; record noise_floor, threshold, contact_signal, peak_taxel_counts.
• Re-measure full_scale_counts and contact_threshold on the physical sensor; decide whether noise_margin: 2.0 holds.
• Confirm the effective threshold separates all four cases; drop the "unverified" note from the sheet.
• Check feat(mcp): gripper_grasp_until_contact, tactile-guided closing #73's grasp_until_contact uses the tare threshold, not the datasheet floor.

Context: #65, Eric's blocking thread on the threshold.