LMDeploy README feature and roadmap lines (read 2026-10-04)

Read 2026-10-04, used to fill the feature-comparison axes for [[niche-lmdeploy-turbomind]]. Fills four axes the record previously left blank, each from a dated roadmap entry or a linked doc: DISAGGREGATED PRE/DECODE - '[2025/06] DeepSeek PD Disaggregation deployment is now supported through integration with DLSlime and Mooncake' (https://github.com/DeepLink-org/DLSlime, https://github.com/kvcache-ai/Mooncake). IMPORTANT NUANCE: this is PD disaggregation reached THROUGH TWO EXTERNAL PROJECTS' integration, not an in-tree scheduler, so the honest reading is 'supported by integration with other projects' rather than 'implemented here' - the same distinction the repo applies to vLLM's third-party connector story. MULTI-MODEL - '[2024/01] Support for multi-model, multi-machine, multi-card inference services' with a dedicated proxy_server doc, and the feature bullet 'Effortless Distribution Server: Leveraging the request distribution service, LMDeploy facilitates an easy and efficient deployment of multi-model services across multiple machines and cards.' MULTIMODAL - '[2023/12] Turbomind supports multimodal input', plus '[2024/07] Support for InternVL2 full-series models, InternLM-XComposer2.5', '[2024/05] Balance vision model when deploying VLMs with multiple GPUs', a Phi-3/Phi-3.5-vision entry in the supported list, and separate VLM inference-pipeline and VLM-serving docs. PREFIX CACHING + KV CACHE QUANT together - 'LMDeploy supports KV Cache Quant, AWQ and Automatic Prefix Caching to be used simultaneously', which is a useful conjunction claim: the three can be composed, not merely individually present. SPECULATIVE DECODING and CHUNKED PREFILL: no mention in the README's feature bullets or roadmap lines, so those two axes remain UNSTATED for this engine and are recorded as such rather than guessed.