feat(rust, python): Streaming joins architecture and Cross join implementation.

[ritchie46]

This sets up the fundamentals of the new out of core query engine. This sets up the architecture for joins and stitches those arbitrary complex query trees together.

I am really happy with this setup, as I think this was the most complicated part of the new engine. Adding the other joins is simply a matter of implementing the proper algorithms.

And as a proof of concept we have a very successful run, streaming through 160GB of data in 15 seconds.

closes #5124

[rjavadi]

I tried the example above with a different filtering condition:

df = (
    lf.join(lf, how="cross")
    .filter((pl.sum_horizontal(pl.col("*")) < 100_000))
)

The explain() output shows that the filtering is inside streaming block, but it fails and causes memory overflow:

--- STREAMING
FILTER [(col("a").sum_horizontal([col("a_right")])) < (100000.0)] FROM
CROSS JOIN:
LEFT PLAN ON: []
  DF ["a"]; PROJECT */1 COLUMNS; SELECTION: "None"
RIGHT PLAN ON: []
  DF ["a"]; PROJECT */1 COLUMNS; SELECTION: "None"
END CROSS JOIN  --- END STREAMING

Does this feature work for window functions such as sum_horizontal?