pub const RGBAF32: Rgba<f32, f32>;
RGBA, 32-bit float per component. This is pretty efficient, since resizing uses f32 internally.
Components are scaled independently (no premultiplication applied)