Support signal processing operators: DFT, STFT, MelWeightMatrix, window functions

[antimora]

Feature description

Implement signal processing operators to improve coverage for ONNX models in Burn:

• Discrete Fourier Transform (DFT) ONNX spec
• Short-Time Fourier Transform (STFT) ONNX spec
• MelWeightMatrix ONNX spec
• HammingWindow ONNX spec
• HannWindow ONNX spec
• BlackmanWindow ONNX spec

These ops are currently missing, which blocks audio feature extraction and spectral model support when importing/exporting ONNX models via Burn.

Feature motivation

Lack of signal processing ops prevents users from running common audio ML pipelines or speech models. Major ONNX models (e.g., for ASR, feature extraction) depend on these ops.

Supporting these ops would:

• Unblock model conversion for audio use-cases
• Enable end-to-end workflows involving audio or spectral models
• Close key compatibility gaps for potential adoption

(Optional) Suggest a Solution

• Template suggested: implement DFT, STFT, MelWeightMatrix, and common window functions per ONNX operator specs.
• Consider using existing Rust crates (e.g., rustfft for FFT-based ops)
• Review PyTorch or NumPy equivalents for reference implementations

Reference issue requesting these ops for Burn-ONNX importer: tracel-ai/burn-onnx#161

Thanks!