[jechol]

This replaces #281, following the suggestion to collect required columns during query construction.

Field references and correlation keys are recorded as expressions and joins are built. The final pass uses this metadata to narrow implicit from_many? selections, without re-scanning expression ASTs. When dependencies cannot be tracked reliably, the original selection is retained.

I've updated the benchmark to compare the upstream base, #281's AST scan, and this implementation with identical dependencies and data.

For the latest Payment query:

• SQL size: 1,532 bytes upstream, 960 with refactor: narrow implicit from_many subquery selections #281, 906 with this implementation.
• Including field collection, query construction, pruning and SQL generation, caller-process reductions are about 1.6% above upstream and 4.6% below refactor: narrow implicit from_many subquery selections #281.
• Results and bind parameters match. Wall times varied between runs, so these measurements do not establish a runtime speedup.

The motivation remains easier-to-read SQL; PostgreSQL already prunes unused projections.

Execution integration is in ash-project/ash_postgres#919.

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies