Scan: exposure lock for nkscan roll scanning

[marcinz606]

Problem

nkscan meters every frame on its own. For color negative it scales each channel so the 99.9th percentile of the frame rectangle lands at 97% of full scale. In the middle of a strip, the brightest area is the orange film base, so the mask comes off before the ADC. At a strip end it is the bare light past the cut, so the mask stays in. Those frames come out 0.4–0.6 D denser in G/B, which is what broke Roll Analysis on the marbella rolls (#1145).

Measured on the scanner, one 6-frame strip:

Frame	Red	Green	Blue
1 (strip end)	82 788	88 160	67 205
2–6	141 635–146 001	351 944–358 393	378 857–389 728

A roll is six strips, so six batches, and the lock has to last across them.

Change

• Scan panel: an Exposure row (Meter Frame… / Unlock) with a hint line, for backends with ScannerCapabilities.exposure_lock (nkscan only). The lock is saved in ScannerSettings, so it survives batches and restarts. A lock metered on another device is ignored.
• Scanner layer: ScanParams.exposures reaches nkscan's scan_frame, which then skips metering. ScanResult.exposures reports the exposures a scan ran at.
• NkscanBackend.meter: meters the whole detected frame with its offset, never the scan window, because the film base in the gap is what the meter should land on. Infrared is included, so the lock also holds for IR and ICE scans.
• Worker: MeterRequest / ScanWorker.run_meter, modelled on prescan. The batch's per-frame offset formula is one function, shared with metering.
• Shortcut registry entry scan_meter_frame (no default key), and a USER_GUIDE bullet.

Test plan

• make all: lint, ty, full pytest (6892 passed)
• New tests: nkscan backend (lock, meter_frame path, fallback, LS-50 superfine), service, worker, settings round-trip, scan sidebar
• nkscan fork hardware test a_frame_scans_at_the_exposures_metered_for_it on the scanner
• On the scanner: Meter Frame… on frame 2, scan 1–6, and check that frame 1's film base matches frames 2–6