The second pass over the CHECK constraint body used a naive
re.sub(r'^[\s\n]*\((.+)\)[\s\n]*$', ...) which, for an expression
like '(x IS NULL OR y IS NULL) AND (x IS NULL OR y IS NULL)', greedily
paired the leading '(' with the trailing ')'. The substitution stripped
both and produced syntactically invalid SQL:

    x IS NULL OR y IS NULL) AND (x IS NULL OR y IS NULL

Replace it with a paren-counting helper that only strips when the
position-zero '(' actually closes at the position-last ')', and that
ignores parens inside single- or double-quoted literals (handling
PostgreSQL's '' / "" doubled-quote escapes).

Fixes sqlalchemy#13157

…re it

Per @CaselIT's review on PR sqlalchemy#13200:

- New util.find_matching_paren(text, start) walks a string finding the
  ')' that matches the '(' at start. Skips single- and double-quoted
  string literals, and treats '' / "" as doubled-quote escapes inside
  their respective contexts (matching PG / SQLite conventions). Returns
  -1 when the opening paren is never closed.
- New util.strip_outer_parens(text) builds on find_matching_paren to
  remove one layer of outer parens iff they wrap the whole stripped
  expression. This is the postgresql/base.py call site.
- sqlite/base.py's existing inline paren-counter (in
  get_check_constraints) is refactored to call find_matching_paren
  instead of its own loop. The sqlite version did the same thing minus
  the doubled-quote-escape handling, which is a strict improvement.

Both helpers covered by unit tests under test/base/test_utils.py
(FindMatchingParenTest + StripOuterParensTest, 21 cases). The postgresql
RegexTest unit tests now import from sqlalchemy.util. The existing
sqlite check-constraint reflection tests (11) still pass with the
refactor.