Add RLVR reward wrapper for IFBench verifiers

[Catnap7]

Summary

Algora bounty: https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4

This PR adds a small train-ready RLVR wrapper around the existing IFBench verifier logic.

It provides:

• rlvr_env.score_response(...) for single prompt/completion reward scoring
• strict and loose evaluation modes matching the benchmark verifier behavior
• binary all rewards and dense partial-credit fraction rewards
• python -m rlvr_env ... to convert prompt/response JSONL files into reward-labeled JSONL
• README usage examples for RLVR loops
• UTF-8 file handling in evaluation_lib.py, which avoids Windows locale decode errors on the included JSONL data
• focused tests for scalar rewards, per-instruction diagnostics, and JSONL output

This is intended as a lightweight adapter for IF-RLVR style training/evaluation loops without changing the underlying instruction verifier implementations.

Verification

Ran locally on Windows with Python 3.12:

.\.venv\Scripts\python -m pytest instructions_test.py tests/test_rlvr_env.py
62 passed, 1 warning

The warning is from the existing syllapy dependency importing pkg_resources; local verification used setuptools<81 to keep that dependency path working.

[Catnap7]

Submitting this PR for the Prime Intellect IF-RLVR/Bench bounty:
https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4

The change adds a train-ready RLVR reward wrapper around IFBench's existing verifiers, with a reproducible CLI and tests. Happy to adjust scope if the expected deliverable differs.