Get matrix multiplies onto the tensor cores by fixing shapes, precision, and alignment, then measure that the cores actually fired. Use when a GEMM or attention kernel runs far below the card's advertised throughput and you suspect it fell back to the CUDA cores.
