DCGM agent integration for real GPU utilization #7

Optionally collect real GPU utilization metrics via NVIDIA DCGM.

Currently EC2 GPU utilization is inferred from CPU/network proxy signals.
DCGM provides actual GPU compute and memory utilization, eliminating
false positives on instances with high GPU but low CPU usage.

• Query DCGM exporter Prometheus endpoint if available
• Fall back to proxy signals when DCGM is not present