feat: add running model path / model precision to perf output

[xieofxie]

Summary

Closes #81.

The perf JSON report did not record which ONNX model was actually run, so users had no way to locate the built/compiled model on disk (issue #81, question 1). This adds the real running model path — the ONNX file ORT actually loads, which can differ from the input model_id/onnx_path when an EPContext (compiled) model is produced or a cached one is reused.

benchmark_info now includes running_model_path:

"benchmark_info": {
    "model_id": "microsoft/resnet-50",
    "running_model_path": "C:\Users\...\.cache\winml\...\model_cpu_ctx.onnx",
    "task": "auto-detected",
    "device": "cpu",
    ...
}

Changes

• session/session.py: track the resolved model path ORT loads via a new running_model_path property (set in compile(); falls back to the input path before compilation).
• session/qairt/qairt_session.py: record the EPContext path the QAIRT session loads.
• models/winml/base.py: expose running_model_path on the model, delegating to the session.
• commands/perf.py: emit running_model_path in benchmark_info, and add it to each per-module instance entry in --module mode.

Tests

• Extended tests/e2e/test_perf_e2e.py::test_benchmark_cpu to assert running_model_path is an existing .onnx file.
• Verified locally (CPU, -m e2e): test_benchmark_cpu and test_module_benchmark_cpu pass; tests/unit/commands/test_perf_cli.py (28) passes.

🤖 Generated with Claude Code
@