Implement multiround benchmarking & enhance LLMService interfaces

[GlebSolovev]

Core Changes ⚡️

• Multiround Benchmarking Support 🥊

  Now it is possible to benchmark multiround proof generation, as is our tradition, with comprehensive error handling and logging.

• LLMService Interface Rework 🕊

  • Simplified error-handling invariants without reducing functionality.
  • Extended proof-generation methods with the ProofGenerationMetadataHolder object, enabling the collection of additional metadata.
  • Improved LLMService constructors for cleaner implementations.
  • Updated GeneratedProof to store raw proof metadata (e.g., statistics returned by the LLMService generation method), facilitating detailed proof analysis—especially useful for benchmarks.
  • Enhanced getter methods for better usability.

Changes to the llm module are thoroughly covered by tests, while the new benchmarking features have been extensively tested manually.

---

Additional Value 🌟

• Enhanced benchmarking result interfaces: redesigned with improved typing for safer and more convenient usage.

• Custom errors: introduced specific error types (IllegalStateError, InvariantFailedError, BenchmarkingError) with wrappers, replacing generic throw Error cases. Refactored the benchmarking module accordingly and ensured custom error declarations are reliable.

• Stricter and clearer error handling: strengthened and documented error handling in the benchmarking framework.

• Utility code improvements: got rid of duplicated utilities, introduced better typings (e.g., for theorem rankers, colorization, LLM iterator).

• Bug Fixes

  • Resolved unawaited promises in affected tests.
  • Fixed error repacking in the OpenAIService.