RDPO support in GRPOTRainer

### Method description

RDPO (https://arxiv.org/abs/2605.13641) is a direct successor to GDPO that addresses two failure modes left unresolved by normalize_then_sum: prompt-level advantage domination under skewed/binary reward distributions, and cross-dimensional reward correlation. It replaces Z-score normalization with Magnitude-Aware Quantile normalization and adds Mahalanobis whitening within each active reward subspace before aggregation.

I am working on a problem with a few different reward directions, and RDPO seems like a better fit than GDPO.

Would a new multi_objective_aggregation="rdpo" mode within GRPOTrainer be welcome? Happy to contribute a prototype PR.

### Open source status

- [x] The method implementation is available
- [ ] The model weights are available
- [ ] The training datasets are available

### Provide useful links for the implementation

_No response_