Refactor reward/verifier setup

[hamishivi]

Bit of a revamp of the way the apply_verifiable_rewards func works:

• We now allow a list of ground truths and answers, instead of just 1.
• We add rewards together in this case
• All verifiers inherit from a basic class that defines the core API.
• We return per-verifier scores along with the added total score, and log out verification scores / rates separately. Note that some verifiers return a continuous value (e.g. the max len sample), so the rate is just counting the times the reward is non-zero.

edit: seems like there is still some bug in logging the rate correctly.... will work it out.

@vwxyzjn, let me know if you want me to make this PR into main, or merge into GRPO-fast and also edit that code for you. I don't mind either way. I need this code for my token control experiments, though!

[vwxyzjn]

Looks good. Let's merge this to grpo packing branch probably.

[hamishivi]

Closing to re-target packing branch.