(EAI-897): Run full NL-to-mongosh benchmark + prompt maxxing

[mongodben]

Jira: https://jira.mongodb.org/browse/EAI-897

Changes

• Add methods for getting Braintrust experiment metadata
  • ended up not using these, but figure that they're good to include anyways as they are tested and functional
• Clean up orchestration for running benchmarks on different prompt types
• Only run benchmarks on a subset of all possible prompt combos (as the total would be unreasonably high, hundreds of experiments and over 100K individual evals)
• Update prompts for optimized performance (prompt maxxing)
• standardize format for experiment names, inspired by URL query params. this makes searching in braintrust for experiments quite easy.
• Add a few more supported models, Gemini 2.5 pro and the GPT 4.1 family

Notes