docs: drop PR reference from SGLang section

docs: note SGLang support is merged to main; recommend vLLM or SGLang for single-node

docs: drop vLLM-specific support clause from intro

docs: note SGLang support

Fix context length: YaRN factor 32->64, max_position_embeddings 131072->262144 (256K, matches FP8/NVFP4 + model card)

initial commit