Allow arbitrary feedback in the reward function other than just accuracy

[laoliu5280]

Ideally, we would like HypoGeniC to be a generic tool that can be applied to any tasks and datasets to generate hypothesis. With the current implementation, it can only support classification tasks because the current reward function takes hypothesis-based inference accuracy and the exploration term.

As an extension, we would like to support arbitrary reward functions, and users of HypoGeniC can choose to use their own reward function / heuristics to evaluate hypothesis.

For example, for text generation tasks, the workflow should be the following:

1. Initialize some hypothesis with a first batch of examples
2. Use the generated hypothesis to perform the text generation task.
3. Compare the generated text and the ground truth text, use a user-defined function to output a quality score of the hypothesis
4. Use the score in the reward function
5. The rest of HypoGeniC should work the same.