[AMD][AgentX] support native ATOM server metrics

[zhuyuhua-v]

Summary

• Add a native ATOM metrics backend that recognizes direct worker metrics (atom:) and Atomesh-normalized metrics (atom_).
• Map ATOM token counters, prefix/LMCache hit rates, and GPU KV-cache usage into the shared AgentX aggregation schema.
• Extend AIPerf dashboard plots to consume ATOM queue, throughput, cache, offload, and preemption metrics while preserving the existing vLLM paths and using token-based units for ATOM LMCache transfers.

Test plan

• python -m pytest -q utils/agentic/aggregation/test_process_agentic_result.py (32 passed)
• python -m py_compile utils/agentic/aggregation/backends/atom.py utils/generate_aiperf_plots.py

[functionstackx]

already merged #2576 which contains this diff