Restructure as publishable crate with CoreML-first depth estimation
Replace candle_onnx with CoreML Swift bridge (default) and ort ONNX
(optional fallback). Port Frame's full API surface: async video pipeline,
multi-format image loader, SBS/TAB/Separate output with MV-HEVC, proper
error types, and model management with correct download URLs.
Feature-gated: coreml (default), onnx (optional), avif/jxl/heic.
Video pipeline now caches the depth model across frames.
Frame can use this as a direct path dep with zero worker.rs changes.