Solver and strength types missing tp_traverse for heap type; repr methods vulnerable to GC-clear

[devdanzin]

Two related type-correctness issues, both primarily relevant under free-threaded Python:

1. Solver and strength missing tp_traverse visiting Py_TYPE(self)

Since Python 3.9, heap types (created via PyType_FromSpec) must visit Py_TYPE(self) in tp_traverse to prevent the type object from being prematurely collected by the GC. The other 4 types in kiwisolver (Variable, Term, Expression, Constraint) correctly do this. Solver and strength do not have tp_traverse at all.

Since neither type holds PyObject* members, adding GC support requires:

• Adding Py_TPFLAGS_HAVE_GC to the type spec flags
• Adding a tp_traverse that visits Py_TYPE(self)
• Changing tp_free to PyObject_GC_Del
• Adding PyObject_GC_UnTrack at the top of tp_dealloc

2. Repr/accessor methods access possibly-cleared members under free-threading

Constraint_repr, Expression_repr/Expression_terms/Expression_value, and Term_repr/Term_variable/Term_value access their PyObject* members (self->expression, self->terms, self->variable) without NULL checks. These members are set to NULL by tp_clear during GC cycle breaking.

Under the GIL this is safe (GC and method calls are serialized). Under free-threading, if GC clears the object while a repr/accessor is running concurrently, the NULL dereference would crash. Adding Py_BEGIN_CRITICAL_SECTION (already used for Variable.context) would protect these accesses.

Low priority — the GC-clear-during-method scenario is very unlikely even under free-threading, but worth noting for completeness.

Found by cext-review-toolkit.