Since we're now (soon ;) letting through float32 arrays in img_as_float, we need to add a fused dtype to all Cython code that consumes doubles to also accept single precision floats.