Fix stubgen emitting invalid syntax for keyword-only dunder parameters

[arose26]

---

Fixes #19315

stubgen produces output that will not parse:

# input
def f1(*, __kw): pass
def f2(x, /, *, __kw): pass

# generated
def f1(*, /, __kw) -> None: ...     # SyntaxError: / must be ahead of *
def f2(x, *, /, __kw) -> None: ...  # SyntaxError

Cause

A parameter whose name starts with two underscores is positional-only by the historical convention, and mypy marks it pos_only on that basis alone. stubgen counts every such argument to decide where the / belongs:

if actually_pos_only_args and arg_.pos_only:
    pos_only_marker_position += 1

The convention does not apply after the keyword-only separator — __kw in def f1(*, __kw) is simply keyword-only — but the argument is still counted, so the marker is inserted past the * that was already emitted.

Change

A keyword-only argument no longer advances the marker position. Genuine positional-only parameters are unaffected, including the double-underscore spelling: def f3(__pos, *, kw) still generates def f3(__pos, /, *, kw).

Tests

testHistoricalPosOnlyParamsWithKeywordOnly covers a keyword-only dunder on its own, mixed with a PEP 570 positional-only parameter, alongside a historical positional-only parameter, and with two of them. It fails before this change with def f1(*, /, __kw) and passes after.

mypy/test/teststubgen.py is otherwise unchanged: 367 passed, with testAttrsClass_semanal failing identically before and after this change on my machine. Self-check on mypy/stubgen.py, ruff and black are clean.