Add PostgreSQL 19 support

[devrimgunduz]

Changelog:

• pgsp_token_types.h: Add a PG19 branch (PG_VERSION_NUM < 200000) for the hand-maintained core-scanner token enum. Values were not guessed: they were extracted from a bison-generated gram.h built from the actual PostgreSQL 19beta1 gram.y (CURRENT_CATALOG et al. stayed the same; FALSE_P/LOCALTIME/LOCALTIMESTAMP/NULL_P/TRUE_P shifted up because new keywords were inserted ahead of them in the grammar). This fixes the "#error This version of PostgeSQL is not supported" and the resulting cascade of ~20 "undeclared identifier" errors in pgsp_json.c (Op, SCONST, IDENT, FCONST, BCONST, XCONST, ICONST, NULL_P, TRUE_P, FALSE_P, CURRENT_*, LOCALTIME[STAMP]).

• pgsp_json.c: Guard the yyextra.escape_string_warning = false; assignment to PG_VERSION_NUM < 190000. The field was removed from core_yy_extra_type in PG19's scanner cleanup, so it no longer exists to assign to; the warning it used to silence isn't emitted by the core scanner in PG19 either, so simply skipping the assignment is correct.

• pgsp_explain.c (report_triggers): PG19 split per-trigger EXPLAIN ANALYZE timing out of the general Instrumentation struct into a new TriggerInstrumentation struct (upstream commit "instrumentation: Separate trigger logic from other uses"). ri_TrigInstrument is now an array of TriggerInstrumentation, exposing firing counts via ->firings and timing via ->instr.total (an instr_time, not a bare double). Added a PG_VERSION_NUM >= 190000 branch that mirrors upstream explain.c's report_triggers(): reads tginstr->firings/tginstr->instr.total instead of instr->ntuples/instr->total, and drops the manual InstrEndLoop() call that's no longer needed/applicable for triggers in PG19. This fixes the "incompatible pointer types", "no member named 'ntuples'", and "invalid operands to binary expression" errors.

• pg_store_plans.c:

  • Add missing #include "access/htup_details.h" and #include "utils/tuplestore.h". PG19's funcapi.h/executor headers no longer pull these in transitively, so heap_form_tuple(), tuplestore_begin_heap(), and tuplestore_putvalues() were showing up as implicit-function-declaration errors even though the functions themselves are unchanged.
  • ShmemInitHash(): PG19 collapsed the old (init_size, max_size) pair into a single nelems argument (4-arg signature instead of 5-arg). Since pg_store_plans already always passed store_size for both arguments, this is a mechanical drop of the duplicate argument, version-gated on PG_VERSION_NUM >= 190000.
  • pgsp_ExecutorStart()/pgsp_ExecutorEnd(): PG19 removed QueryDesc->totaltime entirely and reworked top-of-query instrumentation around QueryDesc->query_instr / query_instr_options, managed by the core executor instead of by extensions calling InstrAlloc()/ InstrEndLoop() by hand (mirrors the same change already made in contrib/pg_stat_statements and contrib/auto_explain upstream). Added a PG_VERSION_NUM >= 190000 branch that requests INSTRUMENT_ALL via query_instr_options before the executor starts, and reads timing/buffer usage back out of queryDesc->query_instr in ExecutorEnd using INSTR_TIME_GET_DOUBLE(). The pre-19 code path (totaltime/InstrAlloc/ InstrEndLoop) is preserved unchanged for older versions.

Hacked by Claude, tested by me. So please double check the patch.

Per: pgdg-packaging/pgdg-rpms#213