Build backend services for metadata, datasets, and training orchestration

[SahilKumar75]

Problem

Frontend-only state is not enough for TuneOS workflows that fetch metadata, index datasets, run training, and manage jobs.

Scope

• Add backend services for Hugging Face and GitHub metadata fetching.
• Add dataset catalog and recommendation endpoints.
• Add training orchestration APIs for LoRA and QLoRA jobs.
• Add error handling, retries, and cache boundaries.

Acceptance Criteria

• Metadata fetching runs through backend services.
• Dataset discovery can be backed by a persistent catalog.
• Training jobs expose status and result endpoints.
• Backend errors surface clear messages in the UI.