Add Support for Mcore-path in Reward Model Environment

## Description

The current reward model (RM) environment does not support advanced features such as mcore-path. By default, the RM environment uses the dtensor path, with dynamic batching and sequence packing disabled.

We should add support for these features to improve the efficiency and flexibility of the reward model. Support for the mcore-path will be considered for implementation after it is enabled for RM training.

If you also need support for the features in the RM environment, please click "👍" on this issue.