Remove pandas as required dependency

[ADBond]

• Get rid of pandas dependency
• Use Pyarrow for table registration/data intechange etc.
• Where we need to do internal calculations on dataframes, can use duckdb

[ADBond]

First bit of work on this: #2883.

[ADBond]

Interesting historical note: #142.

[ADBond]

As part of this I will write up some documentation about how this fits in with the dev infrastructure. But briefly here - we will mark tests that require pandas to be installed to run with a particular marker. This should include only:

• tests that specifically tests pandas-based functionality
• possibly 'lower-priority' backends (postgres, sqlite) - not because they will require pandas to work, but because of the complexity of setting up the testing infrastructure without it. The relevant functionality for these backends should be covered by tests for spark and duckdb anyhow, so it is unlikely that requiring pandas for these tests would mask functionality for these backends that doesn't work without pandas

[ADBond]

Additionally, it would be good to have a set of small focussed tests that don't use pyarrow to check that core functionality works without this. However I think expecting the full test suite to use neither is probably unrealistic.