betto_onnxrt

Reverted out of the monorepo pubspec for now; doc updates for release

Initial development release.

• Native-assets build hook (hook/build.dart) that downloads and stages the ONNX Runtime prebuilt binary (v1.22.0) for macOS, Linux, Windows, Android, and iOS.
• OnnxRuntime — opens the staged ORT library and initialises the OrtApi vtable.
• OnnxSession — generalised FFI inference session supporting arbitrary named inputs and outputs; exposes output tensor shape and element type.
• OnnxTensor — typed multi-dimensional array with named factories for float32, float64, int32, int64, and uint8 element types.
• SessionOptions — thread-pool sizing for intra-op and inter-op parallelism.
• ModelDownloader — SHA-256 verified, crash-safe download of ONNX model files described by a ModelSpec.
• AllowlistProvider — interface for gating which models ModelDownloader is permitted to fetch.
• Added example/magika/ — a standalone Dart CLI tool that detects file types using Google's Magika v3.3 ONNX model. Demonstrates end-to-end use of OnnxRuntime, OnnxSession, and ModelDownloader with a real-world model. Output mirrors the Python Magika --json format. Run with dart run bin/magika.dart <file> or compile with dart build cli from inside example/magika/.