Fix crash on incorrect typing.ClassVar annotation - #1097

…msgspec#1223)

Closes msgspec#1221

The msgspec#1097 fix looks up `ClassVar` on whatever `typing` names in the
module
namespace, but if that attribute lookup itself raises, the error return
is
missed: `structmeta_is_classvar` compares the NULL result against
`typing.ClassVar`, reports "not a ClassVar", and returns 0 with the
exception still pending. On CPython builds with assertions enabled this
trips `assert(!PyErr_Occurred())` in `type_new_init` and terminates the
interpreter.

Return -1 when the lookup fails, matching the error handling already
used
for the `__origin__` lookup just below. The original AttributeError from
msgspec#1096 still propagates exactly as before.

Tests: the existing `test_wrong_classvar` regression is parametrized to
also
cover the subscripted form `typing.ClassVar[int]`, and a new test covers
a
`typing` stand-in whose `ClassVar` attribute raises an arbitrary error,
so
the propagation path itself is pinned rather than only the msgspec#1096 case.

Verified on a locally built CPython 3.13.7 `--with-pydebug`: the
unpatched
code stops the interpreter in `test_wrong_classvar` (exit 134); with the
patch the full unit suite passes (6298 passed, 313 skipped). Focused
tests
also pass on a regular 3.13 build, which is what CI runs.

Changelog entry included under Unreleased.


Tested on Python 3.13 only (debug and regular builds). CI covers 3.10 to
3.15; the change is a two-line error check in version-independent C.

---------

Signed-off-by: Charan Rathore <[email protected]>