fix(runtime): support vLLM 0.26 and SGLang 0.5.16

[ZX-ModelCloud]

Summary

• update vLLM optional-import handling, sampling conversion, token-prompt generation, and config/device discovery for current engine layouts
• support SGLang Engine while retaining the Runtime fallback, including normalized runtime aliases, device IDs, sampling parameters, and token inputs
• accept runtime-supported checkpoint formats, preserve explicit GPU selection, and filter deprecated Evalution constructor arguments