Add support for CMYK -> RGB in decode_jpeg()

[NicolasHug]

Asking for a CYMK jpeg file to be converted to RGB by decode_jpeg() currently results in the following error:

from torchvision.io import read_file, decode_jpeg, ImageReadMode

path = "test/assets/fakedata/logos/cmyk_pytorch.jpg"
decode_jpeg(read_file(path), mode=ImageReadMode.RGB)  # ImageReadMode.UNCHANGED would work

RuntimeError: Unsupported color conversion request

This is an important feature to support because the ImageNet dataset has ~20 CMYK images, which means we currently cannot decode ImageNet files with decode_jpeg(): we have to rely on PIL.

According to this comment in our code there should be a way to support this:

vision/torchvision/csrc/io/image/cpu/decode_jpeg.cpp

Lines 120 to 124 in 9b432d0

	/*
	* Libjpeg does not support converting from CMYK to grayscale etc. There
	* is a way to do this but it involves converting it manually to RGB:
	* https://github.com/tensorflow/tensorflow/blob/86871065265b04e0db8ca360c046421efb2bdeb4/tensorflow/core/lib/jpeg/jpeg_mem.cc#L284-L313
	*/

[tmabraham]

bumping this, had this issue with imagenet just now