Replace torchaudio.load with soundfile.read (bypasses torchcodec requirement)

Remove torchcodec (CUDA-only), use soundfile backend for all torchaudio.load calls

Add torchcodec dependency (required by torchaudio.load in newer versions)

Fix NoneType error when ref_text is not provided

Remove show_api param (not supported in Gradio 6.x)

Remove conflicting gradio==5.9.1 pin, let HF Spaces manage version

Fix gradio_client TypeError: pin gradio==5.9.1, remove sdk_version override

Fix: remove trainer import from model/__init__.py

CPU Basic version: removed spaces.GPU, forced CPU/float32, trimmed deps, whisper-small