OpenVINO GenAI README: speculative decoding, prefix caching and VLM support (read 2026-10-04)

Read 2026-10-04, used to fill the feature-comparison axes for [[openvino-genai]]. PREFIX CACHING is stated in the vendor's own words and unusually precisely: 'Prefix caching that caches fragments of previous generation requests and corresponding KVCache entries internally and uses them in case of repeated query' - note 'fragments', which is a stronger claim than exact-prefix matching and matches how a radix/fragment cache actually behaves. SPECULATIVE DECODING confirmed present (already in the record). MULTIMODAL: 'Visual processing using Vision Language Model (VLMs) - Analyze images/videos with LLaVa, MiniCPM-V and other models' - image AND video. CHUNKED PREFILL and DISAGGREGATED P/D: not present in the README; the record's existing note that continuous batching and prefix caching are actually consumed by OpenVINO Model Server stands, and those two axes are recorded as UNSTATED.