Statistics Precision

[gatesn]

We should follow along this great discussion in DataFusion around how best to handle approximate statistics: apache/datafusion#8078 (comment)

Currently Vortex either has a statistic precisely defined, or it is undefined.

An example of inexact statistics might be that min/max are still valid bounds after filtering an array, but those values are not guaranteed to exist.

[gatesn]

The other question here is about statistics propagation.

Should compute functions propagate stats? Should expressions propagate stats? It's nice to be able to just compute stats for an expression. But does that mean we duplicate all the propagation logic?

Should our compute function "entry points" actually just be the expression and it calls into the VTable directly? But doesn't this create a dependency cycle between vortex-array and vortex-expr?