refactor: separate out pandas and pyarrow usage/imports

[gforsyth]

Today, you can install ibis-framework and generate expressions and compile to any of our backends without installing any of the backend-specific libraries.

However, if you want to actually execute that expression, and get data back, you need pandas and pyarrow installed.
We should separate out and isolate those imports, such that someone could attempt to use only one-or-the-other, or a separate method of pulling data back (like the pycapsule interface), without having those dependencies installed.

We would not be removing them as required imports, because that would make for a bad user-experience when people are getting started, but it would allow for users to explicitly install or uninstall dependencies they don't want, while maintaining other functionality.

xref #10120

[MarcoGorelli]

thanks @gforsyth for considering this!

would it be feasible to aim for something like:

• pip install ibis-framework duckdb lets me run Ibis with DuckDB as engine, without requiring pandas/PyArrow
• pip install ibis-framework[duckdb] brings in pandas+PyArrow too, so that t.to_pyarrow_table works by default and it's a good user experience for people getting started

?

[lostmygithubaccount]

I believe that's the intention -- then in the first case you'll get import errors for things like t.to_pandas(), but most things should work