fix(ai): use atomic operations for refcounting

Replace plain ref_count++/-- with g_atomic_int_inc and
g_atomic_int_dec_and_test for both AIProvider and AISession refcounts.
Prevents data races when ref/unref is called from worker threads.

src/ai/ai_client.h

@@ -22,7 +22,7 @@ typedef struct ai_provider_t
     gchar* org_id;     /* Optional organization ID */
     gchar* project_id; /* Optional project ID (for some providers) */
     GList* models;     /* List of available models (gchar*) */
-    guint ref_count;   /* Reference count */
+    gint32 ref_count;  /* Reference count (atomic) */
 } AIProvider;
 
 /**
@@ -35,7 +35,7 @@ typedef struct ai_session_t
     gchar* model;         /* Model identifier (e.g., "gpt-4", "sonar") */
     gchar* api_key;       /* API key for this session */
     GList* history;       /* Conversation history (GList of AIMessage*) */
-    guint ref_count;      /* Reference count */
+    gint32 ref_count;     /* Reference count (atomic) */
 } AISession;
 
 /* ========================================================================