Rule: detect stopped GPU instances still incurring EBS costs

[maksimov]

Stopped EC2 GPU instances still pay for attached EBS volumes (often large for ML workloads — 500GB+ gp3/io2). These are easy to miss.

Detection:

• Instance state is stopped
• Uptime (time since stop) > 7 days
• Recommend: snapshot and terminate, or restart if still needed

Implementation:

• Add a new rule in internal/analysis/rules.go
• EC2 scanner already discovers stopped instances (state filter includes "stopped")
• EBS cost estimation could use a rough heuristic (e.g., $0.08/GB/mo for gp3) or be left as "investigate"