feat: improve late chunking and optimize pgvector settings

[lsorber]

Changes:

• Add a workaround for Improve late interaction/late chunking context window size #24 to increase the embedder's context size from 512 to a user-definable size.
• Increase the default embedder context size to 1024 tokens (more degrades bge-m3's performance).
• Upgrade llama-cpp-python to the latest version.
• More robust testing of rerankers with Kendall's rank correlation coefficient.
• Optimise pgvector's settings.
• Offer better control of oversampling in hybrid and vector search.
• Upgrade to the PostgreSQL 17.