Make the v3 topic modeling workflow backend-agnostic and add tomotopy support

[sldyns]

Summary

This PR refactors the current pycistopic_v3 topic-modeling workflow so that v3 artifacts are backend-agnostic instead of being tied to Mallet-specific readers and filenames, and adds a tomotopy LDA backend that writes the same artifact bundle.

What changed

• add a shared topic-model abstraction in topic_models.py
• add TopicModelFilenames to centralize the v3 artifact layout
• add backend resolution from the saved parameters file via load_topic_model_backend
• add LDATomotopy to train directly from the binary accessibility matrix and write the standard v3 outputs
• refactor LDAMallet to implement the same backend interface and emit the same artifact bundle
• update the topic_modeling CLI to use a shared run entry point with --backend {mallet,tomotopy}
• keep corpus creation as a Mallet-specific step, while allowing tomotopy to run directly from the matrix plus barcode/region inputs
• update create_anndata, model-stat calculation, plotting, and topic binarization to load outputs through the backend abstraction instead of assuming Mallet-only artifacts
• update log-likelihood/stat handling so backend-specific hyperparameters are interpreted correctly, including learned alpha values from tomotopy
• add tomotopy>=0.14.0 as a dependency

[ghuls]

This pull request will need to be split in multiple commits to be able to review it properly and refrain from rewriting whole functions that e.g. use args.variable instead of the original variable approach.

At first glance it seems like calling run_topic_modeling with the mallet backend will also run the tomotopy topic modeling afterwards.