feat: caching optd stats, 12x speedup on TPC-H SF1

[wangpatrick57]

Summary: Now caching the stat objects used by OptCostModel, meaning we don't need to load data into DataFusion after doing it the first time.

Demo:
12x speedup on TPC-H SF1 compared to not caching stats.

Caching everything except optd stats takes 45.6s total.

Caching everything, including optd stats, takes 3.9s total.

Details:

• This caching is disabled by default to avoid accidentally using stale stats. I added a CLI arg to enable it.
• The main challenge of this PR was making PerTableStats a serializable object for serde.
• The serializability refactor will also help down the line when we want to put statistics in the catalog, since that is fundamentally a serialization problem too. Having Box<dyn ...> would make putting stats in the catalog more difficult.
• This required a significant refactor of how the MostCommonValues and Distribution traits are handled in OptCostModel. Instead of having Box<dyn ...> values in PerColumnStats which store any object that implements these traits, I made PerColumnStats a templated object.
• The one downside of this refactor is that we can no longer have a database which uses different data structures for Distribution (like a t-digest for one column, a histogram for another, etc.). I didn't see this as a big enough reason to not do the refactor because it seems like a rare thing to do. Additionally, if we really needed to do this, we could just make an enum that had both types.

[Gun9niR]

More generally, it can be unique name

[wangpatrick57]

Hm, wdym?

[Gun9niR]

I mean it doesn't have to be a file name. The function's name is a bit confusing, like: file for what? But actually it's just a unique name that can be used as a file name

[Gun9niR]

Rest LGTM