feat: Add a CrossEncoderRerankerHook as a post-processing hook

[jamie-ons]

✨ Summary

This PR adds a new post-processing hook to add a cross encoder reranker. It takes the top retrieved results and reranks them using a cross-encoder model that scores (query, document) pairs jointly, often providing more accurate results than bi-encoder similarity scores (what we currently use in HuggingFaceVectoriser). This adds no extra dependencies and works similarly to the HuggingFaceVectoriser.

I have also included a Demo Notebook to test it out.

The evaluation results with n_results=20 for the reranking and model_name=cross-encoder/ms-marco-MiniLM-L-12-v2:

vectoriser	accuracy	macro_precision	macro_recall	macro_f1
HF	0.770833	0.820833	0.8500	0.816667
HF + CrossEncoderRerankerHook	0.833333	0.900000	0.9000	0.883333
GCP	0.791667	0.845833	0.8625	0.837500
GCP + CrossEncoderRerankerHook	0.833333	0.900000	0.9000	0.883333

📜 Changes Introduced

• Add CrossEncoderRerankerHook to src/classifai/indexers/hooks/default_hooks/postprocessing.py
• Add mps as a torch.device to the HuggingFaceVectoriser and to the hook so huggingface models run faster when developing.
• Add a notebook comparing using the hook on the fake soc dataset.

✅ Checklist

Please confirm you've completed these checks before requesting a review.

• Code passes linting with Ruff
• Security checks pass using Bandit
• API and Unit tests are written and pass using pytest
• Terraform files (if applicable) follow best practices and have been validated (terraform fmt & terraform validate)
• DocStrings follow Google-style and are added as per Pylint recommendations
• Documentation has been updated if needed

🔍 How to Test

• uv sync --all-extras
• run DEMO/reranker_hook.ipynb