feat(perf): add --runtime [ort|openvino] to compare ORT vs OpenVINO

[xieofxie]

What

Adds a --runtime [ort|openvino] flag to winml perf so the same ONNX file can be benchmarked on ONNX Runtime vs OpenVINO Runtime for a side-by-side comparison.

winml perf -m model.onnx --runtime openvino --device gpu
winml perf -m model.onnx --runtime ort --ep cpu   # ORT-native baseline

• Default is ort — existing behavior is unchanged.
• --runtime openvino reads the raw ONNX directly via OpenVINO Runtime (no quantize/optimize/compile build), which is the fair, simple comparison on the same graph. ONNX input only.

How

• OpenVINOSession (session/openvino/openvino_session.py) mirrors the subset of WinMLSession the perf engine uses — compile() / run() / perf() plus io_config / device / ep_name / running_model_path. It reuses get_io_config, load_onnx, PerfStats, and WinMLSession._get_precision, so I/O metadata and reports match the ORT path. No model-specific logic.
• _OpenVINOModel adapter in perf.py exposes the _single surface the benchmark engine reads, so _run_single / _run_benchmark* / reporting are untouched. PerfBenchmark._load_model() branches to it and skips WinMLAutoModel + ORT EP resolution entirely (OpenVINO is independent of ORT's EPs).
• --device maps cpu/gpu/npu/auto → OpenVINO CPU/GPU/NPU/AUTO. compile() fails fast against Core().available_devices with a readable message instead of a raw backend stack trace.
• RuntimeName Literal + RUNTIME_NAMES in constants.py (mirrors CompilerName) — the CLI choice list and the typed config field derive from one source.
• CLI guards: --runtime openvino requires a .onnx input and rejects --module.

Verified locally

• --runtime openvino runs on CPU and GPU end-to-end; latency/throughput populated.
• --monitor works on CPU and GPU (HW utilization via PDH; falls back to NullEPMonitor like most EPs — no OV-specific ep_proof telemetry yet).
• Absent device (NPU here) → friendly error: OpenVINO device 'NPU' ... is not available. OpenVINO sees: ['CPU', 'GPU'].
• New unit tests (tests/unit/session/test_openvino_session.py, gated on importorskip("openvino")) + CLI guard tests; all existing perf tests pass; ruff clean.

Notes / follow-ups

• --ep and quant/optimize flags are intentional no-ops under --runtime openvino (raw ONNX) — documented in the flag help.
• On machines where the WinML registry installs the OpenVINO EP, --runtime ort --device cpu already routes ORT→OpenVINO EP; use --runtime ort --ep cpu for a true ORT-native baseline.
• Possible follow-up: surface EXECUTION_DEVICES in the report so AUTO-mode fallbacks are visible.

Closes #948

🤖 Generated with Claude Code