reliability: implement checkpoint and resume for interrupted runs

[more-shubham]

If processing crashes at image 800/1246, the next run restarts from zero. For large datasets this wastes significant compute time and is a production reliability gap.

Scope:

• Create internal/checkpoint/checkpoint.go managing a checkpoint.json in the output directory
• Track completed file paths with their output hash
• On engine startup, load existing checkpoint and skip already-processed files
• Atomic checkpoint writes (write to temp file, rename) to prevent corruption on crash
• Add -no-resume flag to force full reprocessing

checkpoint.json shape:

{
  "version": 1,
  "started_at": "2026-02-18T10:00:00Z",
  "completed": ["images/photo1.jpg", "images/photo2.png"],
  "total_processed": 800
}

Acceptance Criteria:

• Interrupted run resumes from last checkpoint on restart
• checkpoint.json is never left in a corrupted state
• -no-resume flag bypasses checkpoint entirely
• Checkpoint file excluded from output metrics