Fixed bug when reading BC5S DDS images

[radarhere]

Resolves #7386

In #5501, I added BC5 reading. When I did so, I added the following code to bring SNORM data into the UNORM range.

Pillow/src/libImaging/BcnDecode.c

Lines 121 to 122 in fce23dd

	a0 = (b.a0 + 255) / 2;
	a1 = (b.a1 + 255) / 2;

However, the resulting PNG in our test suite has artifacts - look at the lower middle and upper right of the image.

This was accepted at the time, but #7386 has raised the question again.

Thinking about it now, if the goal is to move from -128 - 127 to 0 to 255, then

a0 = b.a0 + 128;
a1 = b.a1 + 128;

seems like another answer - and indeed, it produces a result without artifacts.