When and Why to Use TMA: A Workload-Grounded Characterization of Tensor Memory Acceleration on NVIDIA Hopper GPUs | ArxivCSExplorer