Implement automated cleanup for uploaded dataset files

[claude]

Problem

Uploaded dataset files stored at {dataDir}/datasets/{userAddress}/{datasetHash} lack an automated cleanup mechanism, leading to indefinite storage accumulation.

Current Behavior

1. Dataset upload (ctrl/task.go:294-333):

   • Files saved to {dataDir}/datasets/{userAddress}/{hash}
   • HF converted format saved to {dataDir}/datasets/{userAddress}/{hash}_hf/
   • No cleanup mechanism targets these directories

2. Existing cleanup (settlement.go:358-378):

   • Only removes task-specific workspace files:
     • {dataDir}/{taskID}/data/
     • {dataDir}/{taskID}/model/
     • {dataDir}/{taskID}/output_model/
   • Does NOT clean uploaded dataset files

3. Unused config field:

   • fileRetentionHours defined in config.go:45-48 but never used in code

Impact

• Storage bloat: Dataset files accumulate indefinitely
• Partial mitigation: Content-addressable storage provides deduplication (same content = same hash = same file)
• Still problematic: Different datasets never get cleaned up

Proposed Solution

Implement cleanup mechanism for uploaded datasets, options:

Option A: Time-based Cleanup (Recommended)

• Use the existing fileRetentionHours config field
• Add cleanup worker to periodically scan {dataDir}/datasets/
• Remove files older than retention period
• Similar to existing task cleanup in settlement.go

Option B: Reference Counting

• Track which tasks reference each dataset hash
• Only delete datasets when no tasks reference them
• More complex but safer

Option C: Combined Approach

• Keep datasets while referenced by any task
• After all referencing tasks are deleted, apply time-based retention

Implementation Checklist

• Choose cleanup strategy (recommend Option A or C)
• Implement cleanup logic in settlement service
• Make fileRetentionHours config actually functional
• Add logging for cleanup operations
• Add metrics for storage usage
• Update documentation

Related Files

• api/fine-tuning/internal/ctrl/task.go (SaveDataset)
• api/fine-tuning/internal/services/settlement.go (existing cleanup)
• api/fine-tuning/config/config.go (fileRetentionHours config)

Context

Identified during PR #337 code review by @Ravenyjh.