[MLX LM] Sampler refactor + a few improvements

[awni]

This is part 1 of a potential second part diff to refactor samplers and logits processors outside of the generate_step function. The idea being that function is getting too fat.. and it will also be much more flexible to experiment with alternative samplers / logits processors without hacking generate_step.

To start this diff:

• Refactors out the sampler and logits processor creation into sampler_utils.py
• Merges the server generation into a single function for stream and completion
• stream_generate is more flexible: returns logits and token as well as text and accepts pre-tokenized inputs
• Use wired limit in stream_generate
• Use stream_generate in the server instead of generate_step
• Put generation in a separate stream to avoid breaking async eval when not using the CPU
• Closes #1032

[awni]

I have some questions that you may or may not have opinions on. Mostly for a follow up PR:

• Thinking about moving make_sampler and make_logits_processor outside of generate_step so that one can pass in any sampler / logit processor to that function. Thoughts?
• Merging stream_generate and generate: yes I want to. generate could easily be a convenience wrapper around stream_generate which does the optional formatting and verbose logging.
• Given the above, I'm considering moving timing/memory reporting into stream_generate to simplify needing to do this in multiple places (it's also easy to do incorrectly) which returns a richer output (maybe a data class with all the fields). I don't want to overcomplicate it.. but that might make it easier to use.

At any rate.. I'll probably put together a PR for some of that soon and we can discuss more concretely. But that's kind of where this is heading in case you have early thoughts on it.