Add Kuzushiji MNIST dataset

[goodhamgupta]

Hi everyone,

Similar to my previous PR(#20), this PR aims to add support for the Kuzushiji MNIST dataset, as mentioned in #11 and #16.

For the KMNIST dataset:

• The train dataset consists of 60k records, with each image having the size 28x28.
• The test dataset consists of 10k records each image having the size 28x28.
• It can be queried as follows:

{train_images, train_labels} = Scidata.KuzushijiMNIST.download
{test_images, test_labels} = Scidata.KuzushijiMNIST.download_test

The API is an exact copy of the mnist.ex file, with the only change being the URLs from which the datasets are downloaded. This PR also adds a quick unit test for the MNIST data loader.

A sample of the images present in the training dataset:

Visualized using:

{{bin_images, img_type, img_size}, train_labels} = Scidata.KuzushijiMNIST.download
images = bin_images |> Nx.from_binary(img_type) |> Nx.reshape(img_size)
images[0..1] |> Nx.to_heatmap() |> IO.inspect

Thanks!

[t-rutten]

Thanks for adding this dataset! It's very cool to see these characters via Nx.heatmap :)

[goodhamgupta]

Thanks for your kind review @t-rutten! I almost forgot about the Nx.heatmap functionality! 😅 I've added a few example images to the PR for reference now! 😄

[t-rutten]

We appreciate your contributions @goodhamgupta :)