PyTorch Devlog Clarifies CUDA Caching Allocator Fragmentation Issues