Fix bicubic stretch interpolation

[homm]

This fixes incorrect result for Bicubic interpolation in im.stretch, discovered by me in comment: #964 (comment)

The error was in line: https://github.com/python-pillow/Pillow/blob/master/libImaging/Antialias.c#L76
Third item was 8*x while should be 8*a*x (see there).

Scale down before and after PR:

Scale up before and after PR:

Source code

# original image
image = Image.open('small.jpg').copy()
image._new(image.im.stretch((117, 77), Image.BICUBIC)).save('small.down.jpg', quality=100)
# make a small copy to enlarge it
image = image.resize((117, 77), Image.ANTIALIAS)
image._new(image.im.stretch((586, 387), Image.BICUBIC)).save('small.up.jpg', quality=100)

[hugovk]

Please can you add tests? This bit of code isn't currently covered by any tests and it's good to improve as we go along.

Apt there was an old comment FIXME: double-check this algorithm here :)

[homm]

I've found another problem! For some reasons, support radius for upscaling were artificially limited to 0.5. Therefore all the filters gave almost identical results for upscaling (same as last picture above).

Without this limitation:
Bilinear:

Bicubic with ɑ = -1.0:

Lanczos:

As you can see, ɑ = -1.0 gives pixelated result for upscaling, so I have changed ɑ to -0.5:

[homm]

@hugovk done.

[homm]

It is better to merge this request before looking at #977

[wiredfool]

Checked this to correspond with wikipedia link.