[AMD] refactor: engine-neutral aiperf plotter + fill sglang panels

[AMD-yanfeiwang]

Summary

• Replace the vLLM-anchored flat alias table in utils/generate_aiperf_plots.py with a semantic Metric enum + per-engine ENGINE_METRICS registry. Panels reference engine-neutral keys; the engine is auto-detected from the metric namespace and the registry resolves each key to the concrete series that engine exports. Adding a backend (e.g. TensorRT-LLM) is now a single new table with no panel changes.
• Support composite metrics via lightweight adapters in the registry (e.g. a ratio of two counters), dispatched transparently by aggregate_timeseries.
• Fill previously-blank sglang panels using upstream (main-branch) metrics only:
  • External/host KV usage from hicache_host_used_tokens / hicache_host_total_tokens (ratio adapter) → drives the External line in the KV Cache Utilization panel.
  • KV offload transfer rate + cumulative from backuped_tokens (GPU→CPU) and load_back_tokens (CPU→GPU). sglang exposes no per-token KV byte size upstream, so these render in tokens; offload panels are now unit-aware (bytes/MB/GB for vLLM, tokens for sglang).
• The per-tier prefix-cache hit-rate split (GPU/External/Combined) and the prefill-source breakdown stackplot remain blank for sglang because they depend on fork-only counters (cache_hit_tokens_l1/l2/l3, cache_miss_tokens) that are not yet merged upstream.

Notes

• vLLM behavior is unchanged: all vLLM metrics stay string mappings, aggregation goes through the same code path, and offload units remain bytes — verified byte-identical output on an existing vLLM-style export.

Test plan

• Re-ran generate_aiperf_plots.py on an sglang result dir; engine auto-detected as sglang, new panels (External KV usage + 3 KV offload panels) populate with real data, fork-dependent panels stay blank.
• Confirmed Metric resolution: CPU_KV_CACHE_USAGE → ratio adapter, KV_OFFLOAD_G2C → sglang:backuped_tokens, KV_OFFLOAD_C2G → sglang:load_back_tokens.
• Sanity-check on a vLLM export in CI/local to reconfirm unchanged rendering.

---

Note

Low Risk
Offline plotting utility only; changes affect visualization labels/units for sglang and refactor metric resolution without touching runtime inference or auth.

Overview
Refactors generate_aiperf_plots.py so panels use a semantic Metric enum and per-engine ENGINE_METRICS registry instead of hard-coded vllm:* names. Engine detection infers vLLM vs sglang from metric namespace prefixes (cached on the export); aggregate_timeseries resolves metrics and can dispatch ratio adapters for composite series (e.g. HiCache host used/total → external KV usage %).

sglang gains upstream-only mappings: external KV line, queue/throughput/preemptions, aggregate prefix hit rate via PREFIX_CACHE_HIT_RATE when hit/query counters are missing, and HiCache offload panels using token units. KV offload panels are unit-aware (MB/s & GB for vLLM, tokens/s & M tokens for sglang). Panels that need fork-only sglang counters (GPU/External prefix split, prefill source stackplot) still render empty. Figure suptitle is generalized to "LLM Server Metrics During Benchmark"; vLLM string mappings are intended to preserve prior vLLM plot behavior.

^{Reviewed by Cursor Bugbot for commit 4f33930. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cquil11]

@AMD-yanfeiwang can you rebase this onto agentx-v0.4 branch please?