[Story] Improve cudf-polars remote IO

This issue summarize some performance issues we're running into with (remote) IO in cudf-polars, and some potential fixes. In this issue, we're primarily focused on remote IO (high latency / high throughput storage systems) from parquet sources, but some of these lessons apply to local IO and non-parquet sources.

## Work Streams

At a high-level, we can break the work into 3 (mostly) independent work streams: optimizing libcudf (and kvikio), getting more information out of polars and using it in cudf-polars, and optimizing cudf-polars.