feat: add after_tables to custom_statements so raw SQL can depend on another table

[jechol]

Motivation

custom_statements had no way to declare "this depends on another table's structure being finalized first." The only rule was "adds always run last, removes always run first." That breaks for a raw-SQL statement that needs to run after some other table is fully set up — e.g. a composite foreign key referencing a unique index that isn't guaranteed to exist yet.

What changed

AshPostgres.Statement gets a new after_tables option (a list of table name strings) so a statement can declare that dependency explicitly:

custom_statements do
  statement :children_parent_composite_fk do
    after_tables ["parents"]
    up "ALTER TABLE children ADD CONSTRAINT children_parent_fk FOREIGN KEY (region_id, parent_id) REFERENCES parents (region_id, id);"
    down "ALTER TABLE children DROP CONSTRAINT children_parent_fk;"
  end
end

Internally this required splitting the single :table_finalized fact into two:

• :table_structure_ready — this table's structural operations (columns, indexes, constraints, etc.) are done.
• :table_finalized — the table is truly done, including any custom_statements declared on it.

A statement's own implicit "wait for my own table" requirement uses the narrower :table_structure_ready, so two statements on the same table never end up requiring each other's completion (a guaranteed cycle if both used the broader fact). A declared after_tables entry uses the broader :table_finalized, so it also waits for the referenced table's own custom statements — not just its structure.

Statements with no after_tables keep the previous default behavior exactly (no declared dependency → best-effort last), so this is fully backwards compatible.

Testing

Builds on the dependency-graph rewrite in the #798 . New tests in test/migration_generator/operation_deps_test.exs cover: a statement's own-table requirement being satisfied by structural ops, an after_tables requirement being satisfied by another table's custom statement (the actual new capability), and two same-table statements not creating a cycle. Full mix test: 858 passed.

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