feat(algorithm): add M2PO for stale-rollout RFT

[primorLee]

Description

Related to #470 and follows the design proposal in #615.

This PR adds M2PO (Second-Moment Trust Policy Optimization) for stable reinforcement fine-tuning with stale rollout data. It implements Algorithm 1 from the final ICLR 2026 paper using Trinity-RFT's policy-loss interface.

M2PO identifies tokens in PPO's active trust-region quadrants, then removes the largest squared log importance-ratio outliers until the remaining second moment is at most tau. Following the paper, masked losses retain the original valid-token denominator and the algorithm does not require a critic, reference model, or KL term.

Changes

• add M2POPolicyLossFn with the paper default m2_threshold=0.04;
• register policy_loss_fn: m2po and the complete algorithm_type: m2po preset;
• add numerical diagnostics for M2 before/after masking, masked fraction, and trust-region fraction;
• test the vectorized mask against a literal Algorithm 1 loop, plus gradient, denominator, registration, empty-mask, and numerical edge cases;
• add an asynchronous Qwen2.5-1.5B GSM8K example;
• document M2PO in the English/Chinese algorithm tables and dataset-oriented example guide.

Paper fidelity

The final paper revision specifies token masking. The authors' earlier public reference code uses adaptive clipping, so this implementation intentionally follows the final published Algorithm 1 and documents that choice.

Validation

• 16 passed, 73 subtests passed for the focused policy-loss and registry suites;
• trainer and explorer YAML files pass full Config.check_and_update() validation;
• targeted Black, isort, flake8, and mypy checks pass;
• git diff --check passes.

The local validation machine has no CUDA GPU, so this PR does not claim a full training-curve reproduction. The included GSM8K configuration is an integration example rather than an exact reproduction of the paper's s=256 setup.

Checklist

• Code has passed all tests (focused tests pass; full GPU/integration CI is pending)
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review