[codex] Update Prime env dataset loading

[sileod]

Summary

• Default the Prime environment loader to reasoning-core/formal-reasoning-env.
• Handle datasets that only expose train by taking eval examples from train after the train slice.
• Filter rows to instruct / few_shot only when a mode column exists.

Validation

• python -B -m py_compile reasoning_core/primeintellect/reasoning_core_env/reasoning_core_env.py
• Installed runtime deps in an isolated Python 3.11 uv venv.
• Confirmed via Hugging Face metadata that reasoning-core/rc1, reasoning-core/symbolic-reasoning-env, and reasoning-core/formal-reasoning-env currently expose only train; direct test loads fail with Bad split: test.
• Smoke tested load_environment(num_train_examples=2, num_eval_examples=2) and confirmed train=2, eval=2.