Designing a TPU and optimizing transformer inference on a $100 FPGA