STEM Benchmarks and verifiers

[gupta-abhay]

• We are going to rely on STEM evals (MMLU-Pro and GPQA-Diamond) as our 3rd subset in Reasoning Gauntlet
• This adds a verifier for these datasets (including extensions for different formats etc.) and also dataloaders

Sample Run name: qwen25-stem-grpo-test-1OAMen
(note: this is with like 1K samples, actual dataset work going on in the background)

[abaheti95]

Thank you! Mostly looks good with minor function renames for future re-use.