Top-down scalar compiler refactor

[carli2]

What

• refactor scalar subquery compile phases to avoid repeated broad raw-AST analysis
• share logical scalar subquery scope preparation between inline and unnest paths
• keep dependent-scalar handling compile-only in fields

Why

• move queryplan compilation closer to the top-down unnesting model from the Neumann paper
• cut compile-time overhead on the real badge/dashboard query without introducing runtime structures

Validation

• python3 run_sql_tests.py tests/32_expr_subselects.yaml
• python3 run_sql_tests.py tests/117_dashboard_unnesting_regressions.yaml
• cold benchmark on the real badge query improved from ~23.69s on master to ~4.5s in this branch