test(e2e): cover streaming usage chunk passthrough across providers

Background

Surfaced from #148 audit (M2). The cross-provider matrix non-stream test asserts `completion.usage.total_tokens` round-trips, but the streaming variant has no `usage` assertion — the test's mock SSE chunks don't include a `usage:` field.

Gap

OpenAI's streaming wire supports `stream_options: {include_usage: true}` which appends a final chunk like:

```
data: {"choices":[],"usage":{"prompt_tokens":N,"completion_tokens":M,"total_tokens":K}}
```

Today the matrix tests don't exercise that path, so a regression where the gateway:

• drops the final usage-only chunk
• mistranslates `usage` field names across providers
• forwards usage on the wrong chunk

would not be caught.

Proposed coverage

Add a streaming case (per provider) that:

1. Caller passes `stream_options: { include_usage: true }`
2. Mock upstream emits the final usage-bearing chunk
3. Test asserts the SDK surfaces `chunk.usage` on the matching chunk

References:

• OpenAI Chat Completions streaming usage: https://platform.openai.com/docs/api-reference/chat/streaming
• Existing matrix file: `tests/e2e/src/cases/cross-provider-matrix-e2e.test.ts`

Priority

P2 — feature-coverage gap, not a regression. Track alongside other matrix-tightening tasks.

Start in tests/e2e/src/cases/cross-provider-matrix-e2e.test.ts and compare the existing non-stream usage assertion with the streaming cases and mock SSE chunks. Add the per-provider stream_options case with a final usage-bearing chunk, then verify that chunk.usage is surfaced on the matching chunk for every provider.