feat(schema): split schema sources so unreadable statements fail alone

[omer-cengel]

Summary

Schema files are now split into statements using JSqlParser's own lexer, and each statement is parsed on its own. A statement the parser cannot read no longer fails the whole file. Real migration histories and pg_dump --schema-only snapshots that contain DO blocks, DROP TYPE, COMMENT ON TYPE, SET search_path TO a, b, identity columns or NOT VALID constraints now load.

Changes

• Schema sources are split at the statement separators the SQL lexer reports. Separators inside strings, quoted identifiers, comments and $$ ... $$ bodies do not split. A tagged dollar-quoted body such as $body$ ... $body$ or $fn$BEGIN ... END$fn$ ends at the next occurrence of its own delimiter. A body that is never closed fails with its line and column.
• psql meta-command lines, such as the \restrict and \unrestrict lines that recent pg_dump releases write, are skipped, and every reported line and column stays a position in the original file.
• A statement the parser cannot read is ignored unless it opens as CREATE/ALTER/DROP TABLE or CREATE/ALTER TYPE. Such an unreadable table or type statement is rejected with the parser's reason at the file line and column of the unexpected token, for example Encountered unexpected token: "DATA" at line 3, column 41.
• An unreadable ALTER TABLE is ignored, without resolving its table, when each of its top-level actions is ADD CONSTRAINT <name>, ADD CHECK/UNIQUE/EXCLUDE/PRIMARY KEY/FOREIGN KEY, or ALTER [COLUMN] <name> ADD GENERATED. This covers the identity form pg_dump writes, as well as EXCLUDE, DEFERRABLE INITIALLY DEFERRED, USING INDEX, NOT VALID and unnamed CHECK constraints. Any other unreadable ALTER TABLE, such as ALTER COLUMN ... SET DATA TYPE, is still rejected.
• Functions and procedures whose body is single-quoted or written with RETURN are now ignored like dollar-quoted ones, and the statements after them are applied.
• DROP TYPE is ignored, so a modeled enum and its columns stay modeled.
• The PostgreSQL guide documents the splitting rules, the ignored and rejected unreadable forms, and the known mis-lex of a backslash-escaped quote in E'...' strings.

Scope and non-goals

• Statements the parser does read are handled exactly as before, including the ALTER TABLE and ALTER TYPE actions that are still rejected, such as SET DEFAULT and OWNER TO.
• DDL inside DO blocks and function bodies is still not modeled, and DROP TYPE ... CASCADE does not remove columns.
• Query parsing, schema namespaces, LIKE and PARTITION OF tables, and COPY data are unchanged.
• The E'it\'s' mis-lex is documented, not fixed.