Retry empty tiny ranged responses without Range header

Accept full-file responses for small ModelScope objects

Use independent HTTP range connections for ModelScope

Resume shards from first unmirrored file

Replace ModelScope SDK downloader with direct HTTP2 ranges

Limit mirror commits to 5 GB

Batch mirror uploads to respect commit limits

Pass large-file size to ModelScope downloader

Use durable full source manifest

Retry paginated manifest and parallelize large downloads

Use ModelScope paginated dataset manifest

Add resumable mirror worker