[GSoC] Better fake data compression script

Context: Since #1661, we have a script replace_fake_images.py which compress all images inside our fake data directory by replacing the random noise images by a uniform color image of the same shape/dtype.

Goal: However, when applied to our fake directory, some files seems to increase size, instead of being smaller (For the list of modified files, see #1634).
For example: imagenet2012/ILSVRC2012_img_train.tar increase by +50KB.
Some JPEG images are also badly compressed by the script (e.g. diabetic_retinopathy_detection/sample/1_left.jpeg).

The goal of this feature request would be:

• Investigate why some archives/image size increase after being compressed.
• Find a way to better compress the images / archives (e.g. by playing with the compression parameters of PIL https://pillow.readthedocs.io/en/5.1.x/handbook/image-file-formats.html#jpeg)

Additionally, other improvement could be done in parallel by different persons:

• Update the script to also compress audio files
• Update the script to also compress video files.

Start with tensorflow_datasets/scripts/replace_fake_images.py and inspect the fake files under tensorflow_datasets/testing/test_data/fake_examples. Run the script against representative archives and JPEGs, comparing input and output sizes while reviewing the PIL compression options. Done means the identified regressions are investigated and image or archive compression is improved; audio and video support are separate possible extensions.