v0.4.1

Inference Pipeline

• `predict()` is now a pure low-level method — eval mode and gradient context are the caller's responsibility
• `inference_on_waveform()` automatically manages eval mode and restores training state via the new `@eval_mode` decorator
• Added `ValueError` for non-1D input in `inference_on_waveform()`
• New `utils/decorators.py` with `eval_mode` decorator using `try/finally` for guaranteed state restoration

Evaluator

• Fixed O(n²) `np.concatenate` — results now accumulate in lists and are concatenated once after the loop
• Replaced `torch.no_grad()` context manager with `@torch.inference_mode()` decorator

Trainer

• Decomposed `train()` into `train_step()`, `val_step()`, and `epoch_step()` — users can now build custom training loops around `epoch_step()`

Module Structure

• Renamed `dtos/` → `schemas/` and split into `items.py`, `predictions.py`, and `types.py`
• `BackboneName` moved to `schemas/types.py` to eliminate circular imports
• Fixed backbone registry typo: `monilenet_` → `mobilenet_`
• Suppressed `FutureWarning` from deprecated `weight_norm` in BEATs backbone

Public API

• Removed internal names (`AudioClassifierConstructor`, `BackboneConstructor`, `BACKBONES`, `POOLING`) from `all`
• `AudioClassifier` and `Backbone` are now the only exported model constructors

CI

• Added `publish.yml` workflow for automated PyPI publishing on GitHub release"