Fix pixel normalization in inference engines; improve DirectML FFI bindings

[Copilot]

Face recognition models (MobileFaceNet etc.) expect pixel input in [-1, 1], but all three engines were normalizing to [0, 1] — causing inference scores to saturate and making threshold changes have no observable effect.

Normalization fix (tflite_engine.dart, onnx_engine.dart, onnx_directml_engine.dart)

// Before — incorrect range [0, 1]
data.add(px.r / 255.0);

// After — correct range [-1, 1] for MobileFaceNet-style models
data.add((px.r - 127.5) / 127.5);

Applied to both NHWC (TFLite) and NCHW (ONNX) preprocessing paths.

DirectML FFI (onnx_directml_ffi.dart)

• Prefers onnxruntime-directml.dll over the standard DLL; falls back gracefully when absent
• Adds proper FFI type definitions and resolves OrtSessionOptions_AppendExecutionProvider_DML at runtime; exposes isDirectMLFunctionAvailable and appendExecutionProviderDML()

DirectML engine (onnx_directml_engine.dart)

Documents the blocker: onnxruntime v1.4.x does not expose the OrtSessionOptions* raw pointer, so the DirectML provider cannot be injected via the high-level Dart API even when the DLL is present. XNNPACK is retained as the best available fallback; DLL detection result is logged.

pubspec.yaml

Added a comment: true DirectML GPU support requires onnxruntime-directml.dll (from Microsoft.ML.OnnxRuntime.DirectML NuGet) placed beside the executable. No separate onnxruntime_directml pub.dev package exists as of v1.4.1.

Note on items already implemented: Test-order shuffling (pairs.shuffle(_random)) and Android GpuDelegateV2 were already in place from a prior session.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.