Make the ONNX runtime path torch-free

[li-lizhe]

Problem

The README advertises:

No PyTorch dependency during inference: it runs directly on ONNX Runtime CPU.

However, onnx_tts_runtime.py does import torch / import torchaudio at module level (lines 12-13), so app_onnx.py and infer_onnx.py fail with ImportError on any machine where PyTorch is not installed. This defeats the purpose of the lightweight ONNX deployment path on resource-constrained devices (ARM64 boards, proot environments — see #73).

Root cause

The only actual torch/torchaudio usage in the ONNX runtime is _load_reference_audio(): audio loading, float32 conversion, resampling, and channel conversion — all only needed for reference-audio voice cloning.

Fix

Replace them with NumPy + soundfile (soundfile is already in requirements.txt):

• torchaudio.load() → soundfile.read(dtype="float32", always_2d=True) — same (channels, time) layout and the same [-1, 1] float normalization convention
• torchaudio.functional.resample() → a pure-NumPy polyphase windowed-sinc resampler (Hann-windowed, 32 taps per side — the same interpolation family torchaudio uses), added as _resample_waveform()
• torch.Tensor.repeat / .mean → np.repeat / np.mean
• soundfile is now an optional import: if voice cloning is requested without it, a clear ImportError with install instructions is raised instead of a crash at module import

Verification

Verified on aarch64 (Kunpeng ARM, openEuler) in an environment with no torch/torchaudio installed:

1. import onnx_tts_runtime now succeeds without PyTorch (previously ImportError).
2. _resample_waveform matches scipy.signal.resample_poly (the same windowed-sinc family torchaudio uses) on band-limited test signals for 44100→24000, 48000→16000, 16000→24000, 22050→16000 — output lengths identical, central-region max abs error < 8e-4 (relative < 8e-4).
3. _load_reference_audio end-to-end on real WAV files: mono 24 kHz passthrough (bit-exact max amplitude), stereo→mono downmix, and 44.1 kHz→24 kHz resampling all produce the expected (1, 1, N) float32 arrays.
4. Identity case (same sample rate) returns the input unchanged.

Fixes #73

[li-lizhe]

Gentle ping on this one — it's been quiet for 27 days since the last activity, so I wanted to check whether there's anything else you'd like me to adjust or add before this can move forward. Happy to make changes if so. Thanks!