Improve memory overhead in PixelsIter

`PixelsIter` stores the previous color values as an array of `Rgb888`, which uses 4 bytes per pixel. The memory overhead could be reduced by using an `u8` array instead.

Using an byte array, where the elements are not aligned to 4 byte boundaries, could cause performance degradation so it might be useful to add a cargo feature to choose between better performance or lower memory consumption.

And the alpha values shouldn't be stored if the image isn't RGBA and a byte array is used. For an array of `u32` pixel values (like `Rgb888` uses) there would be no additional overhead in storing the alpha values, because the high byte is currently unused.