Use exact tokens count when available in server response.

Note that this is not really exacy, as it includes system prompt reminder
and possibly other input data that is dynamically injected.

profiler/src/profiler/TracyLlm.cpp

@@ -825,6 +825,12 @@ bool TracyLlm::OnResponse( const nlohmann::json& json )
 
     if( done )
     {
+        if( json.contains( "usage" ) )
+        {
+            auto& usage = json["usage"];
+            if( usage.contains( "total_tokens" ) ) m_usedCtx = usage["total_tokens"].get<int>();
+        }
+
         bool isTool = false;
         auto& str = back["content"].get_ref<const std::string&>();
         if( !str.starts_with( "<debug>" ) )