Update dask docs

[lhoestq]

• link to blog post for real world use case
• explain how dask parallelism works
• recommend squash history after writing to HF
• add a distributed data processing example
• explain predicate and projection pushdown + example

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[davanstrien]

Looks great! If you don't have a chance I can also add something similar for Polars. I had one question about how Dask does filtering for the downloads. I probably misunderstood something here.

[davanstrien]

Does Dask not still need to download the data to check the values in this column match this filter? From what I understood in the Polars case the predicate push down is usually used for skipping the reading of a column i.e. if you drop it later it doesn't bother to load it and/or doing a filtering step early on. Is Dask directly able to do this before loading?

[lhoestq]

So it will skip the row groups which don't have any row that matches the query using the row group metadata

Then on the remaining row groups it will download the column used for filtering to apply the filter

The other columns will be downloaded or not based on the other operations done on the dataset

[lhoestq]

if you drop it later it doesn't bother to load it and/or doing a filtering step early on. Is Dask directly able to do this before loading?

yes correct ! would be cool to explain that here as well

[davanstrien]

That's super cool!! For some datasets the download time does seem to end up becoming a blocker so this is very neat!

[lhoestq]

thanks for the review ! merging this one for now but lmk if you have more comments