fix: report streaming upstream failures to the request log

[Priyanshu-u07]

stream_upstream cannot raise past its consumer without ending the stream, so it reports upstream failures as data frames. The except clauses in logging_generator_wrapper therefore never fire, and the only detector was
a substring check for Upstream Error — text that appears in a logger.error line, not in any frame the generator emits.

It never matched. Circuit breaker trips, size-limit aborts and provider 500s were all recorded as status_code=200 with no error message, in the inference log and in the Prometheus counters alike, since RequestLogger.log passes the same value to metrics.observe_request.

The failure now travels out of band in a sink dict, the way usage_trackeralready does three lines above. No message text is matched anywhere, so rewording one cannot break it again.

Two things worth a look

The client sees nothing new. json.dumps produces bytes identical to the hand-written frames, so this changes what is logged, not what is sent. There is a test pinning that.

Statuses are no longer flattened to one code. A circuit breaker trip logs 503, a rejected upstream URL 500, and genuine upstream failures 502. That is a judgement call rather than a correctness one, and worth disagreeing with if you would rather a gateway reported 502 for anything upstream.

Closes #373