SGLang patches FP8 cache, image leak

The real story here is not just that SGLang had bugs, but that they lived in edge paths high-performance stacks often miss until users hit them in production.

• –The FP8 issue hit the ragged+paged split in `forward_extend()`, where cached-prefix attention dropped `k_scale`/`v_scale` and quietly degraded outputs
• –That makes FP8 deployments of models like Qwen, DeepSeek-V4, and Gemma 4 more brittle than their BF16 counterparts unless these paths are covered by tests
• –The image-request leak is a classic multimodal cleanup bug: `release_features()` freed pixel tensors but left GPU-resident mrope position tensors behind
• –Silent correctness regressions are worse than crashes because they can poison results while looking “healthy” in observability dashboards
• –If you run SGLang in production, this is a reminder to stress uncommon cache, decode, and vision paths before rolling FP8 or VL traffic broadly