jFloaty: a weird littly JPEG-code and STB-image fork

While it was too cold to dive around here, I was editing some photos and really wished I had a few more bits in the images. Since part of the JPEG compression process is red/green/blue to YUV color conversion, I wondered if decompressing a JPEG with floating-point operations would improve it any.

jFloaty is a fork of STB-image that supports encoding and decoding JPEGs direct to floating-point images, as well as an upgraded PNG encoder with dual 8/16-bit support, and some tools like scaling and dithering. ImageMagick is far more powerful, but at the time didn't support 8-bit to floating-point JPEG decode.

After discussing with libjpeg-turbo's maintainer, dcommander, they it wasn't as simple as I thought - the quality improvements I was seeing were mostly because JPEGs store 10-bit DCT bins for 8-bit images, and they patched libjpeg to support rendering 8-bit JPEGs to a 12-bit surface with a configuration change. This is helpful for software like GIMP and Darktable, who will hopefully pull this in too. I've got a patch up for GIMP and am using it at home - much nicer than an external codec and doing JPEG->PFM->GIMP.

Google has also written a floating-point JPEG codec, jpegli though I didn't see it before making my own. They also have a smarter quantizer than the fixed Q-factor-based quantization tables of most encoders.