Fusing Communication and Compute with New Device API and Copy Engine Collectives in NVIDIA NCCL 2.28

NVIDIA NCCL 2.28 introduces major performance improvements for multi-GPU communication through GPU-initiated networking, device APIs for communication-compute fusion, and copy engine-based collectives. The release enables developers to write custom device kernels that perform network operations directly, offload communication tasks from streaming multiprocessors to dedicated copy engines, and achieve zero-SM operation for certain collectives. New features include the NCCL Inspector profiling plugin for always-on observability, native APIs for AllToAll/Gather/Scatter operations, symmetric kernel group call support, and a flexible environment plugin system. The update also adds CMake build support and enhanced plugin architecture with shared contexts for better cross-datacenter training scenarios.