Implement logic to detect model backend type and route to appropriate inference engine,
enabling seamless switching between llama.cpp and ONNX Runtime backends.

Changes:
- Update model loading useEffect to check backend type
- Route to load_onnx_model for ONNX models, load_model for llama.cpp
- Disable audio capability check for ONNX models (not yet supported)
- Add backend detection in handleSendMessage for inference routing
- Convert image data appropriately for each backend:
  - RGB array for llama.cpp (existing)
  - JPEG bytes for ONNX Runtime (new)
- Call generate_onnx_response for ONNX, generate_response for llama.cpp

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Update list_downloaded_models to include ONNX models by checking for
_onnx_ pattern in directory names and verifying .onnx files exist.
Add frontend model ID normalization to properly match ONNX models from
HuggingFace repos with normalized names (handle slashes, dashes, case).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Integrate ONNX Runtime as an alternative inference backend alongside llama.cpp,
enabling GPU-accelerated inference for vision-language models with platform-specific
execution providers (DirectML, CUDA, CoreML).

Changes:
- Add ONNX Runtime dependencies with platform-specific features (DirectML/CUDA/CoreML)
- Create vlm_onnx.rs module for ONNX inference engine supporting SmolVLM models
- Extend ModelManager with ONNX model download functionality from HuggingFace
- Add Tauri commands for ONNX model operations (download, load, generate)
- Update UI with ONNX Models tab in model selection modal
- Add quantization selector (Q4, Q8, FP16) for ONNX models
- Configure downloads for SmolVLM2-256M-Video-Instruct with correct HF repo structure
  (ONNX files in onnx/ subdirectory, config/tokenizer at root)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Add SmolVLMImageProcessor with 4x4 grid splitting + global image
  (17 frames total, matching SmolVLM's expected input format)
- Add proper prompt expansion with grid tokens
  (<fake_token_around_image>, <row_X_col_Y>, <global-img>)
- Auto-detect layer count from decoder model inputs
- Support multiple EOS tokens (2, <end_of_utterance>, </s>)
- Fix pixel coordinate order to match model expectations
- Add config.json path to model loading for HuggingFace config
- Add test example for VLM inference

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>