Add support for text data & tokenization

[s314cy]

closes #572 and closes #491

add support for text data & tokenization:

• tokenize samples
• load labelled text data in the browser/node
• load unlabelled text data in the browser/node
• lazily load text data in node

this PR includes a rework of the data preprocessing pipeline, which is much more modular and makes it easy to add new preprocessing functions!

it also fixes the CI by:

• making the github actions data cache run-specific
• ensuring the data download script bypasses the gbucket cache
• replacing the example data's archive from a BSD tar to a GNU tar (macOS vs. linux) which caused issues in the CI

[martinjaggi]

should this one be called image?

[martinjaggi]

also mind adding a comment if you will output a stream of token ids?

for LLMs, we can then also support datasets without any label being needed

also let's say where/how people could load different tokenizers (task config or hardcoded either is fine)

[s314cy]

I'll make sure that the PR follows your comments once it's out of the "draft" stage!

[martinjaggi]

very cool, thanks for getting this started!