Migrate from TensorFlow to PyTorch

[nikolaitennant]

Consider migrating the CNN model from TensorFlow/Keras to PyTorch for:

• Better integration with scvi-tools (already PyTorch-based)
• Eliminates the TensorFlow/PyTorch dependency conflict for batch correction
• More active ecosystem for research ML
• Simpler dependency management (single framework)

The core model is a simple 1D CNN — the migration would be straightforward.

Plan

1. Clean up TF version on main (current work)
2. Create pytorch branch from clean main
3. Swap TF model code to PyTorch (~15 lines of model definition)
4. Remove TensorFlow from dependencies, add PyTorch
5. Remove the [batch-correction] optional extra (PyTorch already installed)
6. Tag as v2.0.0 when ready

Note: The published paper (Tennant et al., 2026) states TensorFlow was used. The main branch at v1.0.0 preserves this. The PyTorch migration is a post-publication improvement.