Fast and Compact Tsetlin Machine Inference on CPUs Using Instruction-Level Optimization