Model card: point casual users to the sherpa-onnx ONNX version (callout at the top)

Model card: ONNX version now uses sherpa-onnx; updated speed comparison

Model card: beam vs greedy guidance (accuracy, CPU/GPU timing, when to use which)

Model card: add 'Using the model in PyTorch' (imports, checkpoint format, tensor shapes)

Remove vendored icefall code (install icefall at the pinned commit instead)

Loader uses the user's own icefall checkout (ICEFALL_ROOT); card lists the packages to install

Model card: remove private-repo note

Model card: cleaner layout (summary table, shorter tables, collapsible details)

Model card: how it was trained (code version, settings) and references (papers, code, datasets)

Add license: weights CC BY-SA 4.0, loader code Apache-2.0 (+ vendored icefall Apache-2.0); data license table

Add from_pretrained loader (asr.py) with vendored icefall code; model card with dev+test metrics, datasets and data stats

Bangla Zipformer transducer (pt), average of epochs 23-25

initial commit