Import custom ONNX speech models

[ppXD]

What changed

• allow the local model picker to select .onnx speech models alongside GGML/GGUF
• validate ONNX graphs with the bundled ONNX Runtime and read official sherpa-onnx model_type metadata
• support SenseVoice, Paraformer, sherpa Whisper, and NeMo Parakeet bundles
• automatically discover matching tokens, encoder/decoder/joiner graphs, and external .weights/.data files
• run validation and large-file copying on a blocking worker so the UI stays responsive
• persist imported model size and retain backward compatibility with existing custom-model registries
• document custom ONNX imports in the English and Simplified Chinese READMEs

Why

Custom model import only accepted Whisper GGML/GGUF files even though many downloadable speech models are distributed as sherpa-compatible ONNX bundles. Accepting an arbitrary .onnx file by extension alone would be unsafe because ONNX also contains vision, VAD, embedding, and LLM graphs; Wisp now validates the graph and supports only ASR families it can actually execute.

Validation

• Test First: the initial ONNX tests failed before implementation
• app Rust tests: 28 passed
• root workspace tests: all passed
• new ONNX coverage: metadata mapping, real non-ASR rejection, single/pair/triplet companion discovery, missing-token failure, encoder metadata fallback, external weights
• real-model smoke: downloaded the official 228 MB SenseVoice int8 ONNX + tokens, imported the bundle, initialized the Wisp sherpa recognizer, and completed one inference call
• app clippy with -D warnings
• root fmt/clippy with -D warnings
• Svelte check: 0 errors (one pre-existing missing Node type warning)
• production Vite build passed