fix: drop NaN scores so ranking is deterministic

feat: add INFERENCE_BACKEND provider switch, Space stays as API

Split CPU intake from GPU scoring; 90s reservation

Request 120s GPU duration; quota is checked against requested time

Fix intake reachability: extensionless DICOM, drop header-only pairs

Accept DICOM series (.dcm uploads or .zip) via SimpleITK staging

Accept NRRD/MetaImage volumes via SimpleITK normalize-to-NIfTI

Accept .gz suffix for nii.gz uploads

RADAR ZeroGPU Space