Support SynthPersona QA schema filters

[williamli-15]

Summary

• load current SynthPersona QA metadata such as source, bank_id, family_id, axis, choice labels, baseline flag, and explicit source fields
• add get_qa filters for source, bank_id, and family_id
• switch default train/test split to individual FRQ train rows and shared MCQ test rows
• align BASELINE_PERSONA_ID with the public baseline_assistant persona
• document FRQ/MCQ usage and split_group_id for leakage-aware splits

Tests

• uv run --with pytest pytest tests/test_datasets.py
• uv run python tests/smoke_test.py
• real export smoke with local SynthPersona public preview