I played around with SIMD in Kotlin inside a multi-threaded Mandelbrot generator. My implementation was ~40% slower than the full scalar approach. What happened?