Filter regression in PNG encoding

[mbrubeck]

#298 created a regression that causes some images to be encoded incorrectly as PNG. This following test case demonstrates the problem. The PPM file is correct, but every other row of the PNG file is incorrect.

let white = image::Pixel::from_channels(255,255,255,255);
let black = image::Pixel::from_channels(0,0,0,255);
let img = image::ImageRgba8(image::ImageBuffer::from_fn(100, 100, move |x, y| {
    if (x + y) % 2 == 0 { white } else { black }
}));
img.save(&mut File::create(&Path::new("test.png")).unwrap(), image::PNG);
img.save(&mut File::create(&Path::new("test.ppm")).unwrap(), image::PPM);

Reverting #298 fixes the bug.

Looking at that patch, I think that this line and this line are incorrect, since current[i - bpp] is not the same as orig[i - bpp] in the previous code; it is a value that may have been changed by the preceding for loop.

[mbrubeck]

Another problem is that this code uses -1 as an unsigned integer literal, so range_step_inclusive tries to step by usize::MAX and stops immediately due to overflow.

[mbrubeck]

The issue in the previous comment could have been prevented by rust-lang/rust#5477.

[nwin]

Thanks. Good findings!