fix(translator): preserve usage metrics from trailing choices-empty chunks (#11817)

[HouMinXi]

Fixes #11817. Preserves usage (including cache tokens) from trailing OpenAI chunks with empty choices array in the OpenAI→Claude stream translator. Includes TDD test coverage. ⚠️ base-red inherited: #11449

[diegosouzapw]

Thanks for this — and genuinely good catch. #11883 landed first with a narrower fix
(harvesting usage from the trailing chunk into state.usage), which conflicted with your
branch by the time I got to it. But your PR caught something #11883 actually missed: I
verified directly against #11883's own merged code that when finish_reason arrives
before the trailing usage-only chunk (the normal order for these upstreams), the
client-visible message_delta still went out with {"input_tokens":0,"output_tokens":0}
— state.usage got corrected internally, but too late to matter for the event already
sent. Your pendingClaudeFinishChoice deferral is what actually closes that gap.

I hand-merged your deferral mechanism onto #11883's already-merged trackUsageFromChunk()
extraction (they touched the same lines) as #11933, with your test suite carried over
verbatim. 86/86 tests passing including the existing #5828/#8081/#8162 regression guards.
Full credit for the design and the regression coverage — thank you for tracking this all
the way to the actual emitted event instead of stopping at state.usage. Closing this one
now that it's landed.