Fix: Add explicit Shutdown() call to nanobind cleanup

The nanobind version was missing an explicit Generators::Shutdown() call
during Python module cleanup that the pybind11 version had. This caused
OrtGlobals to be destroyed too late (during C++ static destructors instead
of during Python finalization), leading to use-after-free errors.

The issue manifested as Invalid read errors in valgrind when ONNX Runtime
CUDA provider memory was freed during session cleanup, but then accessed
again when OrtEnv destructor ran.

Changes:
- Added #include "../generators.h" to access Generators::Shutdown()
- Added Generators::Shutdown() call in the atexit cleanup lambda
- Updated comments to clarify the cleanup sequence

This ensures the same cleanup order as pybind11:
1. Python GC releases references
2. Generators::Shutdown() destroys OrtGlobals explicitly
3. Static destructor EnsureShutdown sees null and does nothing

Valgrind results:
- Before: 187 errors (including 2 Invalid reads)
- After: 173 errors (no Invalid reads)
- Matches pybind11 behavior (184 errors, no Invalid reads)

Author: ONNX GenAI Assistant

src/python/python.cpp

@@ -4,6 +4,7 @@
 #include "py_utils.h"
 #include <nanobind/nanobind.h>
 #include "../models/onnxruntime_api.h"
+#include "../generators.h"
 
 namespace nb = nanobind;
 
@@ -77,10 +78,12 @@ NB_MODULE(onnxruntime_genai, m) {
   }, nb::arg("callback").none(), "Set a callback function for log messages. Pass None to clear.");
 
   // Register cleanup function to run before Python shutdown
-  // This ensures our objects are destroyed before LeakChecked counts are checked
+  // This ensures OrtGlobals is destroyed before C++ static destructors run
   auto cleanup = []() {
-    // Force Python GC to run before C++ static destructors
+    // Force Python GC to run first to release Python-held references
     PyGC_Collect();
+    // Explicitly call Shutdown to destroy OrtGlobals before static destructors
+    Generators::Shutdown();
   };
 
   // Use Python's atexit to register cleanup