feat: add IFBench training configuration

[darioandyoshi-tech]

Summary

Adds a minimal prime-rl training config/runbook for the IF-RLVR/Bench bounty candidate.

• Adds configs/ifbench/rl.toml
• Uses the public Prime Intellect Lab env ID shashwat23/ifbench
• Adds configs/ifbench/README.md with run instructions, source links, and validation notes

Validation

python3 - <<'PY'
import tomllib, pathlib
for p in ['configs/ifbench/rl.toml','configs/gsm8k/rl.toml','configs/hendrycks_math/rl.toml']:
    tomllib.loads(pathlib.Path(p).read_text())
    print('ok', p)
PY

Result:

ok configs/ifbench/rl.toml
ok configs/gsm8k/rl.toml
ok configs/hendrycks_math/rl.toml

No heavy GPU training was run locally.

Maintainer question

Please confirm shashwat23/ifbench is the canonical environment ID for the bounty. If a different env ID is preferred, this PR should only need a one-line config update.

---

Note

Low Risk
Low risk: adds new config/docs only and does not change runtime code paths, aside from referencing an external environment ID that may need confirmation.

Overview
Adds a new configs/ifbench/rl.toml training config mirroring existing lightweight RL configs, targeting the slash-qualified Prime Intellect Lab env shashwat23/ifbench (auto-installed by prime-rl).

Includes configs/ifbench/README.md with links, run instructions, and a small TOML-parse validation snippet, plus a note to confirm the canonical environment ID.

^{Reviewed by Cursor Bugbot for commit b3a9e7f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[darioandyoshi-tech]

Hi team, just pinging this for a final review. The IFBench training configuration is ready for merge. Thanks! 🍈