Fix prepared statement cache invalidation on DEALLOCATE

[evkuzin]

When pool_reserve_prepared_statement is enabled, Odyssey tracks which prepared statements have been deployed to each backend connection. However, when a client sends DEALLOCATE ALL, PostgreSQL removes all prepared statements but Odyssey's cache (server->prep_stmts) was not being cleared.

This caused "prepared statement does not exist" errors when:

Client prepares and executes a statement
Client sends DEALLOCATE ALL (e.g., psycopg after ROLLBACK)
Client reuses the same query - Odyssey skips Parse (thinks statement exists), sends only Bind
PostgreSQL returns error since the statement was deallocated
Fix: Detect DEALLOCATE queries and invalidate server->prep_stmts, matching the existing behavior for DISCARD. Handles both simple query protocol and extended query protocol (Parse/Bind).

You can use this script to replicate a bug:

#!/usr/bin/env python3


import sys

try:
    import psycopg
except ImportError:
    print("ERROR: psycopg not installed.")
    sys.exit(1)


def main():
    # Default connection string - connect via Odyssey
    conninfo = sys.argv[1] if len(sys.argv) > 1 else "host=localhost port=5433"
    
    print("=== Odyssey DEALLOCATE ALL Bug Reproducer ===")
    print(f"Connection: {conninfo}")
    print()

    # Reproduce the bug using a temp table (no special permissions needed)
    print("Step 1: Execute prepared statement, DEALLOCATE ALL, then reuse...")
    with psycopg.connect(conninfo, prepare_threshold=0) as conn:
        # Create temp table (exists only for this session)
        conn.execute("CREATE TEMP TABLE _deallocate_test (id int PRIMARY KEY, val text)")
        conn.execute("INSERT INTO _deallocate_test VALUES (1, 'test')")
        conn.commit()
        
        with conn.cursor() as cur:
            # First use - creates prepared statement on backend
            cur.execute("SELECT * FROM _deallocate_test WHERE id = %s", (1,))
            row = cur.fetchone()
            print(f"  First query OK: {row}")

            # Send DEALLOCATE ALL (this is what psycopg does after ROLLBACK)
            print("  Sending DEALLOCATE ALL...")
            conn.execute("DEALLOCATE ALL")

            # Clear psycopg's internal prepared statement cache
            conn._prepared.clear()

            # Reuse the same query - this triggers the bug
            try:
                cur.execute("SELECT * FROM _deallocate_test WHERE id = %s", (1,))
                row = cur.fetchone()
                print(f"  Second query OK: {row}")
                print()
                print("=== RESULT: Bug NOT reproduced (fix is in place) ===")
                return 0
            except psycopg.errors.InvalidSqlStatementName as e:
                print(f"  ERROR: {e}")
                print()
                print("=== RESULT: BUG REPRODUCED! ===")
                print("Odyssey did not clear prep_stmts after DEALLOCATE ALL")
                return 1


if __name__ == "__main__":
    sys.exit(main())