[BUG] Translated OpenAI streams keep the first nonzero usage counters #4263

The translated OpenAI streaming path keeps the first nonzero usage count, even when the provider sends a later cumulative snapshot.

I reproduced this through the actual proxy and installed AnyLLM/OpenAI SDK, using a loopback OpenAI-compatible streaming server. It sends these usage snapshots on successive content chunks:

{"prompt_tokens":10,"completion_tokens":1,"total_tokens":11,"prompt_tokens_details":{"cached_tokens":2}}
{"prompt_tokens":13,"completion_tokens":5,"total_tokens":18,"prompt_tokens_details":{"cached_tokens":9}}

On main fbf7bc3278c11975299b9bd0a2ecf200dfd25a05, /stats records 10 input / 1 output token, and /metrics records 2 cache-read tokens. The client receives both frames intact, so this is accounting after delivery, not a missing upstream event.

The fix records the final 13 / 5 / 9 counts. Usage frames update only fields they actually contain; an absent field leaves the last known value alone, while an explicit zero replaces it. The stateful SSE parser already handles completion-only frames, so the redundant single-chunk parser and its implementation-only tests are removed.

The eight replacement regressions exercise progressive, partial, zero, and absent usage through stream finalization, in Chat and Responses shapes. Six fail before the fix; all eight pass after it. The related streaming/outcome suites pass 71 tests; the shared parser's malformed-event, billed-input, cache-TTL and handler-helper suites pass another 469. The same real HTTP replay reports the corrected counts after the patch.

I'll keep the branch on the fork until I'm below the ten-open-PR limit.