Slice debug track querying has poor performance

Somewhere in the refactorings of dataset (especially when we did create vs createMaterialized) we lost the fact that, if depth is a computed column, it's quite expensive to create it (because it's computed using SQL window functions).

We should really materialize debug tables like this (at least if they're single tables but potentially even for pivoted tables?).

Start by tracing the dataset refactor around create versus createMaterialized and the debug track query path; inspect how computed depth is produced with SQL window functions. Compare single-table and pivoted-table handling, then verify that debug tables avoid repeated depth computation while preserving query results.

Written by the indexing model from the issue text.