Refactor/migrate to pytorch and simplify architecture

[filippogiruzzi]

PR description

** 💥 Breaking change 💥 **
Migrate the Voice Activity Detection project from TensorFlow to a modern PyTorch stack and simplify the codebase. The legacy data_processing / training / inference packages are replaced by a flat vad/ package with four focused modules: data.py, model.py, train.py, inference.py.

Key features:

• PyTorch Resnet1D acoustic model driven by a typed, frozen ModelConfig dataclass.
• uv-based packaging via pyproject.toml (replaces setup.py / requirements.txt), with vad-data, vad-train, vad-inference console scripts.
• Feature pipeline producing 16×65 tensors (5 MFCC + delta + delta² + RMS) from LibriSpeech audio.
• loguru logging across all CLI entrypoints, automatic device selection (CUDA / Apple MPS / CPU), and TensorBoard logging.
• Tooling & CI/CD: ruff lint/format, pytest with coverage in make test, updated Docker images, and refreshed GitHub Actions (CI lint+test, CD Docker build & push).
• Docs: comprehensive open-source README and license alignment to GPL v3.

New behaviour

• Train and export a model with uv run vad-train (checkpoints + TorchScript export, configurable --epochs, --batch-size, --lr, --n-filters, --fc-units).
• Run predictions with uv run vad-inference (optional smoothing, visualization), using a ModelConfig that must match the trained architecture.
• Build the feature dataset with uv run vad-data.
• Device is selected automatically; training metrics are logged to TensorBoard.

Validation tests

• uv run ruff check . — lint clean
• uv run ruff format --check . — formatting clean
• uv run pytest (with coverage) — integration tests pass
• uv sync — package metadata/build valid (GPL-3.0 license)
• CLI: uv run vad-data builds dataset on real LibriSpeech data
• CLI: uv run vad-train trains, exports, and logs to TensorBoard
• CLI: uv run vad-inference loads checkpoint and runs predictions