[build v3] Transitive CTE dependencies are not column-pruned

[shangyian]

In the build_v3 pipeline, CTEs built from transitive dependencies pulled in from dimension links always emit all columns, even when only a subset is used downstream. This is because needed_columns_by_node is only populated for nodes that are directly part of the join path, and not for nodes discovered transitively by collect_refs.

Example

v3.calendar_spine is a dimension node with many columns, referenced inside v3.dim_expanded's SQL:

  -- v3.calendar_spine
  SELECT day_num, label, metadata_col1, metadata_col2, metadata_col3
  FROM v3.src_calendar

  -- v3.dim_expanded
  SELECT a.id, a.some_date, c.day_num
  FROM v3.src_expanded a
  JOIN v3.calendar_spine c ON a.date_key = c.day_num

A user queries a metric with v3.dim_expanded.some_date as a dimension. Only day_num is needed from v3.calendar_spine (as the join key inside v3.dim_expanded's body). But the generated SQL emits all 5 columns:

  -- Current (unoptimized)
  WITH
  v3_calendar_spine AS (
      SELECT day_num, label, metadata_col1, metadata_col2, metadata_col3
      FROM default.v3.src_calendar
  ),
  v3_dim_expanded AS (
      SELECT a.id, a.some_date, c.day_num
      FROM default.v3.src_expanded a
      JOIN v3_calendar_spine c ON a.date_key = c.day_num
  )
  ...

Expected:

  -- Optimized
  WITH
  v3_calendar_spine AS (
      SELECT day_num  -- only what v3_dim_expanded actually references
      FROM default.v3.src_calendar
  ),
  v3_dim_expanded AS (
      SELECT a.id, a.some_date, c.day_num
      FROM default.v3.src_expanded a
      JOIN v3_calendar_spine c ON a.date_key = c.day_num
  )
  ...

Fix Approach

After collect_refs gathers all transitive dep nodes, for each (node, upstream) pair, walk node's SQL AST to find which column references resolve to upstream's table alias and populate needed_columns_by_node with those columns. This analysis must happen before rewrite_table_references renames the aliases, so column refs can still be correlated to source nodes by original table name.