[None][fix] Fix disaggregated cached token usage (feat/bench_x)

[v-shobhit]

Summary

• Cherry-picks the disaggregated usage accounting fix from the original main-targeted PR: [None][fix] Fix disaggregated cached token usage accounting #13620
• Rewrites GEN worker usage with CTX worker prompt token accounting for native disaggregated serving
• Handles both non-streaming responses and streaming SSE usage chunks

Validation

• python3 -m py_compile tensorrt_llm/serve/openai_disagg_service.py tests/unittest/disaggregated/test_openai_disagg_service.py
• python3 -m pytest tests/unittest/disaggregated/test_openai_disagg_service.py -q could not run locally because pytest is not installed in this environment

[dongfengy]

No CI needed in the feature branch. Please merge once the main PR is merged: #13620

[v-shobhit]

Already fixed by Shicheng on 548cac5, closing