Add SGLang as a local inference backend

[kargibora]

We currently support vLLM as the only local GPU inference backend. We should add SGLang as an alternative.Having a second backend gives users flexibility without changing the evaluation workflow and SGLang is also gaining a lot of popularity - which might be handful for future users. What do you think?

[geoalgo]

Indeed, in some cases it would really make sense (i have heard that the constrained decoding is much faster for instance).
One thing though is that it does not seem to be supported in langchain at the moment so we would have to wrap it up ourselves if we want to integrate now without waiting.