Use rai: prefix for top-level RAI fields (validator literal-key match)

Inject 13 RAI fields into croissant.json (incl. all 6 NeurIPS-required: dataLimitations, dataBiases, personalSensitiveInformation, dataUseCases, dataSocialImpact, hasSyntheticData)

Add full experiments/results/ (paper_a + paper_c + rq* + paper_combined; ~397 MB)

Initial benchmark release: 1,539 logged runs + Croissant metadata + claim manifest

initial commit