[GRPO] Reorganize letter counting configs

[wizeng23]

Description

• Changed directory structure in preparation for future PR to add letter counting evaluation.
• Changed model for GRPO letter counting to Deepseek R1 distilled Qwen 1.5B, as reasoning models should have better performance.
• Removed some unnecessary shard_for_eval params for smaller models.
• Fixed broken documentation link.

Related issues

Towards OPE-1122

Before submitting

• This PR only changes documentation. (You can ignore the following checks in that case)
• Did you read the contributor guideline Pull Request guidelines?
• Did you link the issue(s) related to this PR in the section above?
• Did you add / update tests where needed?

[nikg4]

Does this model train w/o errors ? How much slower it is compared to 0.5B model ?

[wizeng23]

There was an unrelated error for model training, which I'm pretty sure didn't exist when i submitted my last PR. After fixing that, it trains. It's a bit slower, at 10 min for 5 steps instead of 6 min. The training speed seems rather variable though; for the 1.5B model, it goes from 5 steps after 10 min to 200 steps after 43 min. It's not just the first step that's slow (like with compilation), but the first couple.