perf(catalog): ERD neighbourhood and MCP describe/sample read the whole schema catalog; filter tables in SQL #1572

Problem

Two paths read the whole schema catalog when they need a few tables:

1. ERD neighbourhood (ErdCatalog.loadNeighbourhood, lib/core/erd/erd_catalog.dart:198-248). It loads all FKs (needed for the graph walk), then for every schema in the neighbourhood runs columnsSql for all columns of that schema and filters in Dart (if (names.contains(r[0])), :230/:238). The PostgreSQL columnsSql has per-column correlated subqueries (unique-index EXISTS, pg_enum string_agg, col_description, pg_get_expr), so the cost grows with the schema, not with the diagram. With catalogRowLimit = 200000 (:19) a large schema is truncated before filtering, so the focused table can lose its columns even though it needs only a few rows.
2. MCP describe_table / sample_rows (McpQueryService._loadCatalog, lib/core/mcp/mcp_query_service.dart:245-270). To find one table it runs three catalog queries over the whole schema (columns, FKs, view columns) on every call. sample_rows uses the result only to validate the table name.

Fix

• Add a table filter to columnsSql / viewColumnsSql: PostgreSQL c.relname = ANY($names) per schema, MySQL table_name IN (...), SQLite m.name IN (...) (literal-escaped as in _lit, or bind parameters where the delegate supports them).
• ERD neighbourhood: walk FKs first, then fetch columns only for the resulting names.
• MCP: validate the name with a cheap existence query (to_regclass / information_schema.tables / sqlite_master), and describe one table with the filtered column/FK queries. Keep the full catalog only for list_tables and the schema:// resource.
• Tests: the generated SQL contains the filter; the neighbourhood of a table in a schema larger than the row limit is complete.

Related: ERD epic #1166 and #1167 (same catalog SQL).