Add a scoring page that wraps viame score on desktop and web

[mattdawkins]

• New top-level Scoring tab (desktop between Pipeline and Settings, web after Models) and a Score bulk action in the library.
• Score any number of sequences in one run: per row pick the computed and truth annotations (current, annotation set / revision on web, earlier annotation file or browsed file on desktop, or another dataset with the same footage). Pairs are exported as computed/seq_N.csv + truth/seq_N.csv and scored by one viame score folder-mode job.
• Results stored as auxiliary/scoring_<ts>.json on the first computed dataset; "Previous runs" lists every run the user can read; "Reuse setup" restores a run's sequences and parameters.
• Tabs: summary metrics (detection, localization, segmentation, keypoint, length, MOT/HOTA/KWANT), per-class table, confusion matrix (click a cell to filter errors), interactive PR/ROC curves (hover for threshold, consolidated vs per class, click a point then "Use as confidence filter"), threshold sweep with recommended per-class filters, per-object error browser with per-frame chart.
• All viame score options exposed in a parameters dialog (IoU, confidence, box/polygon matching, per-class, top-class, aux confidence, tracking, keypoint tolerance, sweep steps, filter estimator, default label, class synonyms).
• Web: POST dive_rpc/score, GET dive_scoring, GET/DELETE dive_dataset/:id/scoring[/:resultId], GET dive_dataset/:id/scoring_sources; CSV export gains a set param. Desktop: RunScoring job on the CPU queue.
• Requires the matching VIAME change (VIAME/VIAME tool-applets 11f12f500) for --match-mode, --output-matches and inlined sweep curves.
• The shared viewer is untouched. Verified end to end on desktop; the web worker path is unit-tested only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YEefheVfnAJsDhdisaBmis

[BryonLewis]

Most changes are kept relatively separate from affecting other code.

Below are some changes I've made to this branch I think to imporve user experiance and during my testing.

• Source linking for scoring runs: Scoring UI updates so pairs can point at annotation sets, revisions, or on-disk sources, with viewer navigation into those annotations.
• Web dataset picker: Dialog-based dataset picker on web for choosing comparison datasets in scoring.
• PDF export cancel: Cancel support in the PDF preview flow for scoring exports.
• Sample scoring dataset script: samples/scripts/scoring/generate_dataset.py and README to generate test data for the scoring workflow.
• Annotation JSON import fixes: Better handling of .dive.json files (import path fix in the VIAME serializer and improved JSON annotation detection on import).
• Picker remembers last folder: Add-sequence / dataset picker opens to the last folder used.
• VIAME virtualenv fix: Clears VIRTUAL_ENV and adjusts PATH in viame_config.py so VIAME's embedded Python is not pointed at the DIVE venv.