Problem

WorkflowNode surfaced a nested workflow's result as "the last yielded
event with non-nil Output" — which is iteration-order dependent. This
diverged from adk-python and was wrong for two common shapes:

• Fan-out with multiple terminals: whichever terminal finished last
  by wall-clock won, so the result was non-deterministic across runs.
• Silent terminal: if the terminal node produced no output but an
  earlier node did, the intermediate node's output was returned as if it
  were the workflow's result.
  The existing linear tests passed only because the single terminal happens
  to emit last (TestNestedWorkflow_MultipleOutputs, despite its name, is
  a linear Chain).

Solution

Select the output from the workflow's terminal node(s) — nodes with
no outgoing edges — computed from graph topology rather than event
arrival order, mirroring adk-python's _set_ctx_output_or_interrupts:

• exactly one terminal output → forward it;
• more than one distinct terminal produced output → ErrMultipleOutputs
  (a workflow must have a single output identity; use a JoinNode to
  merge branches);
• zero → no output.
  Outputs are keyed by terminal node name (last write wins), matching
  adk-python's node_outputs dict, so re-activating a single terminal
  counts as one output rather than an error. Terminal output is read via
  childEventOutput, so a terminal that carries its result as model text
  (MessageAsOutput) is handled too.

A new graph.terminalNodeNames() helper computes the terminal set
(mirrors adk-python's graph._terminal_node_names).