feat: add trajectory-level batching, loss aggregation, and dynamic batching support

[lqzxt]

Summary

This PR introduces trajectory-aware training support by batching samples at the trajectory level, aggregating loss across entire trajectories, and adding dynamic batching to better handle variable-length trajectories. These changes ensure that all multi-turn completions from the same trajectory are processed within the same forward/backward pass while improving training efficiency and reducing memory fragmentation.

Changes

• implement trajectory-level mini-batch construction to keep all completions from the same trajectory in the same batch
• refactor loss computation to support trajectory-level loss aggregation
• add dynamic batching for variable-length trajectories to improve GPU memory utilization and reduce fragmentation