[Fix] test_streams_tokens_then_eos passes when the engine fails #146

On torchnative 0.1.0b4 with mps (#84) every engine test failed, but test_server.py::test_streams_tokens_then_eos passed: BaseModel.chat turns an engine exception into an ERROR: message frame, and the test only checks that some text arrived before <EOS>.

Done when: the test fails if the streamed text is an error message (asserted by checking it does not contain the chat error marker), and passes on a working engine.