Upgrade MLX framework, and refactor create_generator to use stream_generate

[neilmehta24]

MLX package upgrades

mlx==0.21.0
mlx-lm==0.20.1
mlx-vlm==0.1.1

Each of these brings new features and improvements to the engine.

MLX LM upgrade

• Switch over to using the stream generate API instead of generate step. stream generate is a more stable API, and also includes a wired limit setter. This addresses the issues described in Set wired limit before starting generation #40
• Start using the new mlx_lm sampler class, since passing temp min_p and other sampling params into generate_step is deprecated. We will use mlx_lm default sampler method, and pass in user provided sampling params.
• Repetition penalty params are now required to be passed in through the logits processor, so make that refactor.

MLX VLM upgrade

• This upgrade adds support for two new models, Florence 2 and Molmo.

• Florence requires as an input into the language model, the token that was generated during the last evaluation. Use the mlx_lm custom sampling capability to store the most recently sampled token as part of vision_model_wrapper. Note that this model requires trusting remote code.
  

• Molmo required no special code refactors. Note that using the Molmo model may use a lot of memory, so I have been testing by resizing smaller than usual. Note that this model requires trusting remote code.
  

generate.py and demo.py updates

• Start requiring named arguments for create_generator. Assign defaults for all of the named arguments, so that the caller doesn't have to provide defaults for arguments that may not be informed about. The generate_args option is removed, and now all of the configurable parameters are part of the method signature. The generate_args dict for mlx_lm is built up during create_generator
• Since we have moved to the stream_generate API, there are a couple of refactors for passing arguments to mlx_lm, and handling return values during generation. This PR Stop strings with stream generate #45 fixes stop string support, it isn't handled correctly in this PR

[neilmehta24]

[None] is a more correct implementation. This will flatten the array, instead of just taking the first element

[mattjcly]

When I initially was testing to set this to 1000x1000, if I dropped it down to 512 models like qwen2vl would start mis-recognizing text from screenshots of my screen.

Why the change? If for certain models can we custom resize just for those models?

[neilmehta24]

I didn't mean to commit this change. I reverted this.