Add smoothness training pipeline with STFT/adversarial/vocoder-consistency losses

- inflect_nano/train_smooth.py: enhanced training with multi-res STFT loss,
  adversarial mel discriminator, vocoder consistency loss, deeper residual
  postnet, and cosine LR schedule
- preprocess_dataset.py: convert HF datasets, local dirs, or LJSpeech CSVs
  to the durations.jsonl format needed by training
- inference.py: add --smooth-prosody, --mel-smooth-sigma, --lowpass-hz flags
  for zero-cost inference-time quality improvements
- test_inference.py: smoke tests for model loading and synthesis
- colab_smooth_finetune.ipynb: Colab notebook for T4 GPU fine-tuning
- requirements.txt: add numba, scipy, datasets

requirements.txt

@@ -5,3 +5,6 @@ numpy
 g2p_en
 transformers
 gradio
+numba
+scipy
+datasets