Better design pattern for data_weight synchronization

The ready event in neighbor_allreduce dst_weight makes sure the data_weight computation is done before communication, as Pytorch CUDA stream is not synchronized with our CUDA stream.

Start by tracing the neighbor_allreduce dst_weight ready event and how data_weight computation interacts with PyTorch and the project’s CUDA streams. Determine a synchronization design that preserves the required ordering before communication; done means the synchronization pattern is defined and the relevant behavior is validated.

Written by the indexing model from the issue text.