fix(peaks): peak marks follow chemical-shift reference edits

[Limdongcheng]

Summary

PeakMark stored its x as a finished-spectrum snapshot, so editing a Reference step shifted the whole ppm axis while every existing mark stayed at the old coordinate — visibly detached from its peak.

Marks now store the uncalibrated position (finished x minus the pipeline's reference offset at pick time) and every reader adds the current offset back through PeakSet::resolve(offset), fed by the new Dataset::peak_reference_offset_ppm() — the pipeline's single semantic outlet chemical_shift_reference_offset_ppm, previously used only by CRAFT. The calibration lives once, in the pipeline, so marks follow any Reference edit — including undo/redo — without being rewritten. Both write paths (manual picks and detection) subtract the current offset symmetrically; de-duplication compares in uncalibrated space. Persisted mark coordinates change meaning; the format stays v1 with no compatibility reader, per repository policy. y deliberately remains a pick-time intensity snapshot: reference edits never change intensities.

Validation

• cargo pr-check passes on current main (fmt, source sizes, dependency policy, default-frontends build, clippy -D warnings, reference-backend test suite).
• Regression tests: mark → reference edit → resolved ppm follows the shifted trace and its default label; picks on an already-referenced spectrum round-trip through the uncalibrated store.