arrow2 erased refcounted clones performance issues

[teh-cmc]

That one is very bad, this is obviously something we do all over the place.

Background:

• Basically all arrow Arrays are Buffers when you reach the actual physical type (which can be quite deep because all our components are compound types etc)
• Buffer is literally an Arc over some bytes, so cheap to clone
• But then we carry around erased arrow arrays, i.e. trait objects: Box<dyn Array>
• Trait objects cannot implement Clone because they literally dont know their size/layout at compile time
• But there's a hack for that
• All of the above is what I mean when I say "erased refcounted clone"

See #1745 for detailed benchmarks.

This is yet another example where fixing the issue once in DataCell and then using DataCell everywhere would be nice...

[teh-cmc]

Written down some detailed thoughts about this in #1745 (comment).

The TL;DR of it being: wrap the body of a DataCell in an Arc, and enjoy a life with less problems.

The core issue is the same as in #1738: see #1738 (comment).