Replace EOL PyPy 3.9/3.10 with PyPy 3.11 in CI matrix

[srchilukoori]

Summary

• Replace EOL pypy-3.9/pypy-3.10 with pypy-3.11 in CI matrix
• Skip gssapi extra on PyPy — the C extension is CPython-only
• Remove gssapi_require from tests_require so extras are composable
• Extract gssapi-specific tests to test_auth_gssapi.py with pytest.importorskip

Fixes #606

Problem

Two issues broke PyPy CI:

1. cryptography >= 47.0.0 dropped pre-built wheels for PyPy < 3.11. Without a wheel, pip builds from source and fails because PyO3 0.28.3 requires PyPy >= 3.11. Dependency chain: keyring → SecretStorage → cryptography.

2. gssapi is CPython-only. On PyPy 3.9 a cached pre-built wheel happened to work. On PyPy 3.11 it builds from source and produces a .so referencing _PyGC_FINALIZED — a CPython internal not present in PyPy:

ImportError: gssapi/raw/creds.pypy311-pp73-x86_64-linux-gnu.so: undefined symbol: _PyGC_FINALIZED

Root cause

The CI install step and setup.py both assumed gssapi is universally installable. tests_require bundled gssapi_require, so even pip install .[tests] pulled CPython-only C extensions on PyPy.

Fix

Layer	Change
setup.py	Remove gssapi_require from tests_require — .[tests] should not pull CPython-only deps
ci.yml	Install .[tests,gssapi] on CPython, .[tests] on PyPy
test_client.py	Guard kerberos/gssapi imports with try/except; use pytest.param(..., marks=requires_gssapi) for the shared test_authentication_fail_retry
test_auth_gssapi.py (new)	Pure gssapi tests with pytest.importorskip("gssapi") — entire module skips cleanly on PyPy

On CPython with .[tests,gssapi], all tests run as before. On PyPy with .[tests], gssapi tests skip with a clear reason.

Test plan

• CI build (pypy-3.11, latest, ~=1.4.0) passes with gssapi tests skipped

• All CPython matrix jobs pass with full gssapi test coverage

• Verified locally: test_client.py — 46 passed, 2 skipped; test_auth_gssapi.py — 1 skipped (module-level importorskip)

• Yes, this is an AI-assisted PR.

[hashhar]

Thanks for the PR. One question which I maybe didn't understand properly.

[hashhar]

test failure looks related. The approach looks good to me though.

[srchilukoori]

@hashhar , @azawlocki-sbdt Please merge the PR when you get a minute. thank you.