v0.7.2

• feat(models): add Qwen3.8 27B generation profiles and stabilize speculative streaming
• perf(inference): persist SGLang kernel caches and tune Qwen3.8 hardware profiles
• fix(models): pin Vidore revisions and scope the ColQwen revision correctly
• feat(storage): add native Alibaba OSS payload storage and ACK/RRSA Helm support
• fix(helm): harden KEDA scale-to-zero, ACK storage selection, and cleanup behavior
• feat(tooling): publish the OSS development tooling and mise task surface
• fix(server): report engine startup crashes accurately