fix(perf): poll the right adapter under --device gpu/--monitor (#445)

[xieofxie]

Fixes #445 — winml perf --monitor --device gpu now shows GPU utilization instead of silently falling back to NPU.

Summary

• Adapter routing. PdhPoller / HWMonitor now accept device= (npu | gpu | cpu | auto) and ep_name=. Per-device queries go through build_npu_query (Compute engine) or build_gpu_query (3D engine).
• ORT-first LUID resolution. Added resolve_adapter_luid(device_kind, ep_name=None) in sysinfo/pdh_adapters.py. It walks ort.get_ep_devices(), filters by EP name + device type, reads OrtHardwareDevice.metadata["LUID"], and converts the decimal LUID to the PDH 0xHHHHHHHH_0xHHHHHHHH form via (v >> 32, v & 0xFFFFFFFF). ORT-resolved LUIDs are sanity-checked against enumerate_adapters() — if PDH doesn't enumerate the LUID (some drivers register the adapter only for autoEP), we skip it and fall through to the PDH-fingerprint helpers (discover_npu_luid, discover_gpu_luid) instead of raising later in build_adapter_query. Same fallback applies when ORT isn't installed, when get_ep_devices() raises, when no matching ep_device exists, when LUID metadata is missing, or when an individual ep_device entry is malformed (each is wrapped in try/except so one bad entry can't break resolution).
• Graceful build-query degradation. PdhPoller.start() wraps build_npu_query / build_gpu_query in a local try/except ValueError. If a resolved LUID still proves unusable, it clears device_kind / adapter_luid and degrades to CPU/RAM-only — CPU/RAM counters still register and the benchmark continues.
• perf.py plumbing. _run_benchmark_monitored (HF), _run_onnx_benchmark (ONNX direct), and _perf_modules (per-module) all pass the resolved device + normalized EP name into HWMonitor. --monitor help text updated to "Show live hardware utilization chart for the benchmarked device (NPU, GPU, or CPU)".
• HWMonitor.to_dict() schema. Always emits an npu block (zeros when monitoring GPU — back-compat with existing consumers/tests). When device_kind == "gpu" a parallel gpu block is added. New top-level keys device_kind and adapter_luid.
• Live chart labels follow what's actually polled. Centralised adapter_label(device_kind) helper in hw_monitor.py. LiveMonitorDisplay accepts device_kind= (threaded through from _run_monitored_loop as hw.device_kind); HWLiveDisplay derives the label dynamically from self._hw.device_kind. So auto resolving to GPU now labels "GPU", and CPU-only mode no longer falsely says "NPU".
• CPU-only display drops the adapter row entirely. When no adapter is polled (--device cpu or auto resolving to nothing), LiveMonitorDisplay / HWLiveDisplay skip the green plot, drop the swatch from the chart legend, and collapse the status row to CPU: ... | Sys Mem: ... (no adapter cell, no Device Mem cell). The post-benchmark display_console_report in perf.py likewise prints CPU/RAM only — no zeroed NPU: 0.0% avg, 0.0% peak.

Verification

• Targeted suite: tests/unit/session/test_ep_monitor.py (84 passed, +15 new) plus test_perf_cli.py and test_perf_module.py (26 passed).
• Full unit suite green: 4665 passed, 73 skipped, 2 xfailed.
• ruff check clean on every modified file.
• Smoke-checked on a Snapdragon X laptop after winml.register_execution_providers(ort=True):

  GPU (no ep_name)  -> 0x00000000_0x000121CF   # Adreno X1-85 via DML and QNN
  GPU (ep=DML)      -> 0x00000000_0x000121CF
  GPU (ep=QNN)      -> 0x00000000_0x000121CF
  NPU (ep=QNN)      -> 0x00000000_0x00012C89   # authoritative, not just fingerprinted
  NPU (no ep_name)  -> 0x00000000_0x00012C89