[#52] Implementation of format reward

[yilinjz]

Implementation of format reward as described in Issue #52.

Current training logs revealed three specific structural failures correlated with the reward drop:

1. The model fails to close tags.
2. The model attempts to call tools (e.g., search_code) in the final turn instead of providing an answer.
3. The model hallucinates the raw <|endoftext|> string.

Thus, we introduce a positive-only format reward to guide the model back to structural stability.