feat(audience): add locate / line / transcription qualification examples

[RapidPoseidon]

Summary

The Rapidata dashboard now offers Locate / Line / Transcription audience types alongside the existing Classify and Compare (app-frontend#2342, with the LineExampleTruth multi-box backend support landing in rapidata-backend#3992 + rapidata-backend#4051). The Python SDK only had typed helpers for Classify and Compare — adding the missing three so the SDK matches the dashboard's surface and the dashboard's "Create it with Python" snippet on those forms can point at a real public method.

What's new

Three typed methods on RapidataAudience, each calling through to a matching method on AudienceExampleHandler:

audience.add_locate_example(
    instruction="Click the dog's face",
    truth=[Box(x_min=0.3, y_min=0.2, x_max=0.55, y_max=0.45)],
    datapoint="https://assets.rapidata.ai/dog.jpg",
)

audience.add_line_example(
    instruction="Trace the horizon line",
    datapoint="https://assets.rapidata.ai/landscape.jpg",
    truth=[Box(x_min=0.0, y_min=0.45, x_max=1.0, y_max=0.55)],  # optional
)

audience.add_transcription_example(
    instruction="Select the words actually spoken",
    sentence="the quick brown fox jumps over the lazy dog",
    truth=[1, 2, 3],  # word indices
    datapoint="https://assets.rapidata.ai/fox.mp3",
)

All three:

• mirror the keyword-arg shape of the existing add_classification_example / add_compare_example (same context, media_context, explanation, settings),
• compute a sensible randomCorrectProbability (calculate_boxes_coverage(truth) for Locate, same for Line when boxes are given, 0.1 flat for Line when not, len(correct) / len(total) for Transcription),
• run through _try_start_recruiting after the example is added.

Refactors / housekeeping

• validation.rapids.box: extracted calculate_boxes_coverage from the private RapidsManager._calculate_boxes_coverage so both the rapid and audience layers can reuse the union-area sweep-line. The old method on RapidsManager is kept as a thin alias so any external callers (and any docs that reference it) keep working.
• IExampleTruthLineExampleTruth codegen now carries boundingBoxes / requiredPrecision / requiredCompleteness (matching prod after rapidata-backend#4051). The local generator-cli can't materialise Java in this dev environment, so the model file was hand-patched in the same shape the generator would produce; the next CI regen will overwrite it idempotently.
• Schema files refreshed from prod via bash openapi/generate-schema.sh.

Docs

• docs/audiences.md: added a method-picker table for all five example types and a runnable snippet showing the three new helpers. Updated the existing 'settings' note to reflect that all add_*_example helpers accept settings, not just classify/compare.

Test plan

• pyright src/rapidata/rapidata_client — no new errors above main's baseline (the 3 remaining errors are pre-existing missing-import warnings for pandas / pydantic in this isolated environment).
• Smoke: audience.add_locate_example(...) against a staging audience, verify the resulting rapid carries a LocateBoxTruth with the expected boxes and requiredPrecision / requiredCompleteness propagated.
• Smoke: audience.add_line_example(..., truth=None) and audience.add_line_example(..., truth=[Box(...)]) and verify both round-trip — the first produces an empty LineTruth, the second a BoundingBoxTruth (per the rapidata-backend#4051 mapping).
• Smoke: audience.add_transcription_example(...) and verify correctWords reach the rapid as TranscriptionTruth.correctWords.

🔗 Session: session-2de1ad62