Add standalone irreps tensor product environment

[DengZhiyuan-math]

Summary

• add a standalone O(3) irreps tensor-product environment with deterministic local generation
• add exact symbolic scoring with whitespace normalization and strict parity/term validation
• document the environment and include focused plus generic environment smoke coverage

This environment does not depend on reasoning-gym or e3nn.

Validation

• uv run pytest tests/test_irreps_tensor_product_env.py -vv
• CHANGED_ENVS=irreps_tensor_product uv run pytest tests/test_envs.py -vv -k "pyproject or readme or env" (local mock OpenAI-compatible endpoint)
• uv run ruff check environments/irreps_tensor_product tests/test_irreps_tensor_product_env.py tests/test_envs.py
• uv run ruff format --check environments/irreps_tensor_product tests/test_irreps_tensor_product_env.py tests/test_envs.py
• direct vf-eval smoke against the local mock endpoint: exact reward 1.0

Note

Add standalone irreps tensor product decomposition environment

• Adds a new irreps_tensor_product environment (irreps_tensor_product.py) that generates O(3) irreps tensor product decomposition problems and scores model answers by exact canonical match within <answer> tags.
• Problems are generated deterministically from a seed and index; answers are normalized to a canonical l{e|o} + ... format with whitespace tolerance before comparison.
• Config validation raises ValueError at load time for invalid parameter ranges (e.g. max_l < min_l, both parities disabled).
• Updates test helpers in tests/test_envs.py to shell-quote paths and support custom eval API endpoints via VF_EVAL_API_BASE_URL and VF_EVAL_API_KEY_VAR.

^{Macroscope summarized 351b554.}