This PR factors out model name and minihf_infer inference parameters into config files. This lets us do things like

• Load a generator without an evaluator, for manual looming
• Load generators and evaluators that aren't LoRA finetunes of the same base model
• Initialize the minihf UI with different parameters per config file

I plan to extend the config to handle params for the weave.py CLI as well, but that'll rely on my other PR's code: #7

The default behavior of python minihf_infer.py should be the same as it is on main, or you can try python minihf.infer -c configs/gpt2.config to use the UI with GPT2 locally and no evaluator model.