fix(pretty-print): preserve PostgreSQL syntax in five printer edge cases

[edjubert]

What

Five independent output defects in the printer.
Four of them made the beta round-trip guard refuse the statement (Normalized ASTs differ after formatting), so the user silently got no formatting at all.
The fifth is a layout regression.

1. ALTER TABLE ... ADD COLUMN IF NOT EXISTS

missing_ok was never emitted, so the statement came back without IF NOT EXISTS and the guard rejected it.

ALTER TABLE s.t
	ADD COLUMN IF NOT EXISTS row_id TEXT,
	ADD COLUMN IF NOT EXISTS designation_source TEXT;

Dropping the flag would turn an idempotent migration into one that fails on its second run, so the guard was right to refuse - the fix is to emit it.

2. UNIQUE NULLS NOT DISTINCT on a table constraint

nulls_not_distinct was emitted for CREATE INDEX but not for a table constraint, where the qualifier sits between the keyword and the column list:

CREATE TABLE s.expense_types (
	expense_type_legacy_id TEXT,
	legacy_branch_code TEXT,
	UNIQUE NULLS NOT DISTINCT (expense_type_legacy_id, legacy_branch_code)
);

Losing it would let two NULL rows coexist where the schema forbids it.

3. A boolean expression as the argument of a boolean test

AND, OR and NOT bind more loosely than the postfix IS test, so (a OR b) IS TRUE was printed as a or b is true, which reparses as a OR (b IS TRUE).
The argument is now parenthesised when - and only when - it is a BoolExpr.
Every other argument kind binds tighter and needs nothing.

WHERE (units.has_history OR units.has_shares OR units.has_calls) IS TRUE
  AND (units.is_active AND units.is_visible) IS NOT FALSE

4. DISTINCT ON parentheses

The opening parenthesis emitted a soft line but the closing one did not, so a broken list left the ) glued to its last element:

-- before
select distinct on (
  a,
  b)
  a,
  b

-- after
select distinct on (a, b)
  a,
  b

The clause now also gets its own group, so it can stay on one line while the target list breaks - the usual shape of a DISTINCT ON query.
This changes several existing snapshots (groupingsets, memoize, select_distinct_on, subselect, tsrf), all in that direction.

5. typeCase on a qualified type name

CAST(t.id AS public.object_id) went through the plain identifier path, so typeCase did not apply to it while it applied to bigint.
A new emit_type_identifier_maybe_quoted emits a TYPE_IDENT token for names that can be written unquoted, and falls back to the verbatim identifier for names that need quotes - "MyType" and "MYTYPE" are two different types in PostgreSQL, so recasing them would change which type is referenced.

Tests

Three new fixtures (qualified_type_cast, distinct_on_long_list, unique_constraint_nulls_not_distinct), snapshots for two fixtures that had none (alter_table_add_column_if_not_exists, boolean_test_bool_expr), and two unit tests on the new emitter covering the quoted/unquoted split.