Use outlines's controlled generation

[yujonglee]

https://normal-computing.github.io/outlines/reference/controlled_generation.html
This can be useful for more complex output, such as multiple scoring.

Main blockers are:

• Use jinja based prompting #38. It does not support Python 3.9
• This only support some of models https://github.com/normal-computing/outlines/tree/main/outlines/model
• Slow(?) Haven't tried, but might be.