Connect to the BC6 decoding from Python

[radarhere]

Helps python-pillow#6449

Two changes here.

1. Connect to the BC6 decoding from Python

After this PR, with your code as well,

from PIL import Image
im = Image.open("source.dds")
im.save("out.png")

gives

3. Change bc6_finalize for unsigned data.

As per the comment in the code, the half_to_float function is taken from https://gist.github.com/rygorous/2144712. It is half_to_float_fast5. Except, that code is for converting FP16 into FP32. FP16 has 1 sign bit, 5 exponent bits and then 10 bits for the fraction. Unsigned BC6H has 5 exponent bits and then 11 bits for the fraction. So instead of scaling it down from 65535 to 31743, I'm just scaling it down to 255.

[ShadelessFox]

Your image looks faded because of truncation to uint8. I don't think that's the result we want to achieve.

I believe the BC6 decoder should accept an additional parameter, gamma, just like the BC7_UNORM_SRGB variant.

Then, the result of bc6_finalize should be gamma-corrected, normalized to be in the range 0..255 (since it's an HDR container and its values may go beyond 0..1), and only then scaled down to 255.

[ShadelessFox]

Fixed in 5087356