Decouple dynamic_set_t from the table name collector and the serializer

[trueqbit]

Summary

dynamic_set_t, the SET clause assembled at runtime with dynamic_set(storage), serializes each assignment as it is pushed back, since its assignments are of different types and can't be kept as a tuple. At the same time it collects the tables named on the assignments' left-hand side, which the serializer needs later for the UPDATE. Before this PR:

• it kept a table_name_collector as a member;
• push_back() was defined in ast/crud/set.h;
• so this AST node header included table_name_collector.h, forward-declared iterate_ast(), and relied on serialize() being found later through ADL.

With this PR:

• The node keeps only the result: dynamic_set_t::table_names, which clear() resets.
• push_back() moves out of class into implementations/dynamic_set_definitions.h. It uses a local collector, and that header includes what its body uses: ast_iterator.h, table_name_collector.h and statement_serializer.h. The header is listed in the interface_definitions.h manifest, whose comment now mentions AST nodes alongside schema and storage classes.
• collect_table_names() of a dynamic SET returns set.table_names.

ast/crud/set.h no longer depends on the collector, the AST traversal or the serializer. That leaves table_name_collector.h used only by serialization code, which prepares for grouping the serializer's helpers later.

TODO.md notes the remaining coupling: the node's type carries the serializer context, and it serializes eagerly. Keeping the assignments type-erased instead would let a dynamic SET be serialized and bound like any other node.

Test plan

• New: tests/statement_serializer_tests/ast/set.cpp checks that push_back() collects the table names and clear() resets them.
• Existing coverage of the dynamic SET: the serialization tests in the same file, tests/statement_serializer_tests/statements/update.cpp (UPDATE with a dynamic SET), and tests/storage_non_crud_tests.cpp.

🤖 Generated with Claude Code