Refactor ``ConfigGenerator``

[vijusudhi]

Currently, ConfigGenerator does the minimal setting with one IC config, (LoRA + FFT) finetuning and corrseponding evaluations.

TODOs:

• Extend this to different possible configurations
  For example:
  • IC experiments can be conducted on different prompt templates, consider using Dspy?
  • Zero-shot, few-shot
  • Fine tuning experiments can be extended to different PEFT techniques
  • For PEFT fine tuning, different hyperparams can be tried
• How to extend the Train and Eval Configurations?
  Maybe use Hydra and hierarchical configurations for this?