Document the pyctcdecode numpy<2.0.0 pin workaround, and the new graceful LM-degradation behavior

Add architecture/language-head/performance diagrams

Rewrite model card: detailed architecture, language-head story, SVG diagrams, corrected speed claim

Upload config.yaml with huggingface_hub

Upload lm_5gram.bin with huggingface_hub

Upload model.pt with huggingface_hub

Upload bpe.vocab with huggingface_hub

Upload bpe.model with huggingface_hub

Upload config.yaml with huggingface_hub

Upload README.md with huggingface_hub

initial commit