Evaluate inclusive metrics from borrowed file statistics

[unikdahal]

Part of #3343.

What's the feature are you trying to implement?

InclusiveMetricsEvaluator decides whether a data file might contain rows matching a predicate. It only reads the file's record count and its per-column value, null and NaN counts and lower/upper bounds, yet its entry point requires a complete DataFile. Statistics that live outside a DataFile — for example whole-file statistics carried with a scan task for execution-time pruning — cannot be evaluated without building one.

Proposal: introduce a crate-internal view that borrows exactly the statistics the evaluator reads, evaluate through it, and keep the existing DataFile entry point as a thin conversion.

Expected behaviour:

• No change for existing callers: evaluating a DataFile gives the same results as today.
• A record count known to be zero still excludes the file unless empty files are included.
• An unknown record count does not by itself exclude a file; the available bounds and counts may still prune it.
• No public API change (pub(crate) only).

This is a small refactor that lets later #3343 work prune whole files from task-level statistics without reconstructing DataFiles.

Willingness to contribute

I can contribute to this feature independently; an implementation with tests is ready.