MLXLanguageModel loads model on every respond()

[noorbhatia]

MLXLanguageModel calls loadModel() on every respond() and streamResponse() invocation, causing the model to be repeatedly reloaded from disk. This leads to unnecessary overhead and poor performance.

[mattt]

Yikes. That was definitely not the intention. I could've sworn that we had a single call to loadModel on first response. But if that's not the case, then we should definitely fix that.