Provide guideline and starting script for contributors working on changing model specs

[chenmoneygithub]

We need guidelines for our contributors working on model specs, such as compilation defaults. What we need to do:

• provide a guideline on what should be tested, and what to write in the PR description. For compilation defaults, we may also ask for comparison across a few choices.
• Provide starting scripts, including classification/MLM/Causal LM/.. so that contributors could just plug in the model to check the changed part.

[Cyber-Machine]

@chenmoneygithub I think we should also make a template for issues related to working on model specs so issue tracking is consistent in the future.

Scripts already available, PR expectations, datasets used, and a link to the guideline can be highlighted.

[chenmoneygithub]

@Cyber-Machine Thanks for sharing your thoughts! Personally I am not a fan of issue template, it's a bit hard to strike a balance between useful and verbose. But yes, it could be useful to set a guideline on what should be included in the PR description, especially when the PR is a breaking change.