llama.cpp adds ModernBERT reranker support

llama.cpp has released an update adding support for classifier_pooling in ModernBERT rerankers, enabling native semantic reranking within the library. The implementation reads pooling types from model configurations and supports 'cls' and 'mean' methods, with a fallback to mean for unspecified cases. This enhancement allows developers to integrate high-performance reranking models directly into local retrieval-augmented generation (RAG) pipelines without relying on external services or complex setups.