A Forward Pass for One Token: Where the FLOPs Go

Update (Oct 29, 2026): while measuring Part 3 I found a second hidden slowdown: cubing numbers with x**3 is very slow in NumPy when the numbers are negative. It inflated the "other" row below. I fixed