[experiment] Layout Reader

[gatesn]

What feels like the 27th time I've explored this space, I think I might finally be getting somewhere.

This design pulls out essentially a scan engine. Layouts are actually just one way take serialized arrays and construct a ScanPlan, but in theory we could build a ScanPlan by hand or by any other means.

A ScanPlan node can accept push-down of various operations:

• try_push_expr - apply an expression over the plan node. This is the closest to our current layout readers that can take an expression, intercept some of it, pass some down to a child, and so on.

This plan can then be used to answer different types of questions:

1. prepare_read - evaluate a row range of the plan to produce a resulting array.
2. prepare_evidence - construct zero or more evidence providers that contribute evidence to a predicate. For example, zone maps are an evidence provider that emits definitely_false, but not definitely_true results, and with a much lower cost than a filter evaluation.
3. prepare_aggregate - return the requested aggregate partial state. This allows zone maps or other layouts to answer aggregate push-down queries.
4. prepare_stats - return a set of approximate aggregate partials for the given aggregates.

[more description to come]