feat(api): validate sequence annotation completeness on transition to seq_annotation_done

Context

The invariant "every object in a sequence annotation must be classified before the sequence reaches seq_annotation_done" is currently enforced only in the frontend:

• frontend/src/utils/annotation/progressUtils.ts — isAnnotationComplete() requires bboxes.every(hasUserAnnotations) plus a missed-smoke review.
• frontend/src/utils/annotation/sequenceUtils.ts — hasUserAnnotations(): a bbox is classified iff (is_smoke=true and smoke_type set) or false_positive_types is non-empty.
• frontend/src/components/sequence-annotation/AnnotationHeader.tsx — save button disabled unless complete or marked unsure.

The API accepts any processing_stage with any payload — nothing re-validates completeness server-side.

Why it matters

The frontend is not the only writer: push_sequence_annotations.py, update_annotation_stage.py (make update-stage-remote), import.py, and any raw API client can move an annotation into seq_annotation_done.

The detection-review pipeline (B) then trusts the stage blindly and silently mislabels incomplete data: collect_annotation_bboxes() in pull_sequence_annotations.py treats an object with is_smoke=false and empty false_positive_types — i.e. an unreviewed object — as a false positive and exports it as fp_unlabeled. If the FiftyOne reviewer doesn't catch it, apply_fiftyone_review.py deletes the remote detection annotations and recreates them with that wrong label at stage annotated. This is the same class of classification-fidelity bug as #148/#150.

Proposal

When a sequence annotation is created or updated with processing_stage = seq_annotation_done and is_unsure = false, validate that every entry in annotation.sequences_bbox satisfies:

• is_smoke = true and smoke_type is set, or
• false_positive_types is non-empty.

Return 422 with a field-level error otherwise. Natural home: the existing validation layer in app/schemas/annotation_validation.py / the sequence-annotation endpoints (which already derive has_smoke / has_false_positives server-side, so there is precedent for annotation-content logic in the API).

Caveats

1. Validate on the transition (POST/PATCH carrying that stage). A stage-only PATCH must validate the stored annotation payload.
2. Audit existing rows first: any legacy incomplete annotation would make stage-only updates start failing (update_annotation_stage.py, the push script). Probably desirable failures, but the count should be known before enabling.

Start with app/schemas/annotation_validation.py and the sequence-annotation endpoints, then inspect how POST/PATCH requests derive has_smoke and has_false_positives. Audit existing rows before enabling the check, including stage-only updates that use the stored payload. Done means transitions to seq_annotation_done with is_unsure=false reject incomplete sequence objects with a 422 field-level error, while complete payloads continue to work.