fix: remove torch dependency from ONNX runtime

[z2z23n0]

Summary

• replace torch/torchaudio usage in onnx_tts_runtime.py with soundfile + NumPy
• keep reference-audio output shape compatible with the ONNX codec encoder: (1, channels, samples)
• add soundfile>=0.12.0 to pyproject.toml because the ONNX runtime now imports it directly

Closes #73.

Root Cause

The ONNX runtime path still imported PyTorch because reference-audio loading used torchaudio.load, torch.float32, and torchaudio.functional.resample. That meant app_onnx.py could fail at import time in lightweight ONNX-only environments without PyTorch installed.

Validation

• Created a fresh Python 3.12 venv with only lightweight ONNX-path dependencies: numpy soundfile sentencepiece onnxruntime
• Verified torch_available=False and torchaudio_available=False
• Verified import onnx_tts_runtime succeeds without PyTorch installed
• Verified _load_reference_audio() loads a temporary mono WAV, resamples it, duplicates channels, and returns float32 shape (1, 2, 4000)
• Ran python -m py_compile onnx_tts_runtime.py
• Ran git diff --check