vLLM v0.25.0: Model Runner V2 as Default and PagedAttention Removal

The latest release of vLLM (v0.25.0) introduces significant changes that affect local AI operators running dense models. The new Model Runner V2 is now the default execution path, bringing support for EVS, real-time embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding compatible with full CUDA graphs. Additionally, PagedAttention has been removed as the legacy implementation is no longer necessary with MRv2 becoming the standard backend. These changes can impact performance and model compatibility on consumer hardware.