Memory use trait / methods / macros

In #19599 we were discussing adding a `heap_size()` method to `Statistics`, which requires adding a similar method to all of the structs it references. The idea occurred that there are already crates to handle this sort of thing, namely:
- Our own [HeapSize](https://github.com/apache/arrow-rs/blob/b93fa52e47a83dff130709a6e80a4be5017dbf09/parquet/src/file/metadata/memory.rs#L33-L42) trait in `arrow-rs`
- [deepsize](https://crates.io/crates/deepsize) which is now unmaintained and it's [maintained but less widely used fork](https://github.com/nhtyy/deepsize2)

It seems that if we're going to be implementing this for multiple structs it would be nice to at least have a derivable trait and use that trait everywhere.

One choice is between using a 3rd party crate or our own trait. Another choice is what the semantics we want are, the most confusing one perhaps being ref counted pointers (`Arc`, `Rc`, etc.).

Having thought about it a bit I think we should go with `HeapSize` and buff that trait up to provide a macro, implement it for the standard library container types, etc.

That said I'm opening the issue to collect community feedback if anyone has opinions.

Start by reading issue #19599 and the existing HeapSize trait in arrow-rs, then compare the deepsize and deepsize2 approaches. Clarify whether DataFusion should use its own trait or a third-party crate, including semantics for Arc and Rc. Done means the community has agreed on the trait design and implementation scope; no DataFusion file or test is named here.

Written by the indexing model from the issue text.