Various minor fixes

[dmontagu]

• Use the _build_graph method for building the agent graph in the Agent class, rather than a direct call to the agent graph function
• Fix the fact that we were closing the run_span right away when graph runs were auto-instrumented.
• Tweak the implementation of the Graph.run loop so that we don't modify the value of start_node, which is a bit confusing, and allows us to remove a branch from the conditional. The remaining code is closer to what I'd expect a user to make use of (though I still want to simplify further, as demonstrated in Add GraphRun object to make use of next more ergonomic #833)