feat(grammar): PERFORM, EXECUTE, RETURN QUERY, INTO targets, DO and CALL

[sebsnyk]

Stacked PR, converted to draft. This sits fourth in the order #371 -> #372 -> #373 -> #374 and depends on #373. GitHub cannot base a fork PR on another fork branch, so the diff below includes the predecessors; only the last commit (90e976f) is this change. It is not ready for review until #371 to #373 land and it is rebased; closing it until then is a reasonable call.

Problem

The last group of statements that real PL/pgSQL bodies use every few lines: perform, execute ... into ... using, return next / return query, select ... into strict, insert ... returning ... into, and the do $$ ... $$ block that migrations wrap them in. The sample in #236 has RETURNING * INTO session_record as its one remaining unparsed construct after #371 to #373; with this PR it parses without an ERROR node. Against real sources:

source	ERROR nodes with #373	with this PR
PostgreSQL src/test/regress/sql/plpgsql.sql	363	252
pg_partman sql/functions/check_default.sql	9	6
utPLSQL source/core/ut_utils.pkb	233	232

pg_partman's remaining six are @extschema@ templating placeholders and SET search_path = ... in the function header (#364, separate PR). utPLSQL still needs Oracle's header shape and package bodies.

Solution

A new grammar/statements/dynamic-sql.js:

• perform_statement: PERFORM [DISTINCT] select_expression [from ...], reusing the select tail so joins, where, order by and limit come for free.
• execute_statement: EXECUTE [IMMEDIATE] expr [INTO [STRICT] target, ...] [USING [IN | OUT | IN OUT] arg, ...]. IMMEDIATE and the argument modes are PL/SQL's spelling of the same statement.
• _function_return extended with RETURN NEXT expr and RETURN QUERY (query | EXECUTE ...).
• SELECT ... INTO [STRICT] ... in _select_statement, and RETURNING ... INTO [STRICT] target, ... on returning, so insert, update and delete all take it. Both are optional keywords in an existing position; existing trees are unchanged.
• OPEN cursor [SCROLL] FOR EXECUTE ... on the open_statement from feat(grammar): declaration attributes and cursors in procedural bodies #373.
• do_statement: DO [LANGUAGE x] $$ [<<label>>] [DECLARE ...] BEGIN ... [EXCEPTION ...] END $$ or a string body, in statement so it works at the top level.
• call_statement: CALL procedure(args), also in statement.

Notes

• tree-sitter test: 538 corpus tests on feat(grammar): declaration attributes and cursors in procedural bodies #373, 548 with this PR (6 new in test/corpus/execute.txt, 2 in do.txt, 2 in call.txt), plus the 2 highlight tests. No ERROR/MISSING; scripts/test-keywords.sh passes.
• Parse table: STATE_COUNT 31202 → 31151 (-0.2%), but LARGE_STATE_COUNT 5346 → 5512 and parser.c 42.7 MB → 43.2 MB (+1.2% on top of feat(grammar): declaration attributes and cursors in procedural bodies #373; +3.8% over main). DO and CALL sit in statement, so unlike the body-only rules in the earlier PRs they reach every statement context; that is where the large-state growth comes from. If that cost is unwelcome, the two could be limited to program.
• Not included: COMMIT / ROLLBACK inside a procedure body. transaction shares the body statement list since feat(grammar): procedural control flow in function and procedure bodies #371, so a COMMIT statement inside a body would be ambiguous with the one that ends a BEGIN ... COMMIT transaction; that needs a decision about how the two rules should relate.