SDK version is painfully out-of-date #186

The library hasn't been updated in over a year. There are about two dozen configs that are now missing, including:

• enable_thinking to enable/disable thinking
• reasoning budgests
• over a dozen model loading options, such as CPU MoE offload, spec decoding, autofit, etc.

I'm currently working on bringing all of the missing configs over from the lmstudio-js SDK. There were a couple of things I had questions on, however.

For example, there's a llama.autoFit and an mlx.autoFit and (looking again) yuzu.autoFit and vllm.autoFit. for server->client that's easy (just consider both). what do I do on client->server? Send all of them?