Feat/fix vertex chunk parsing instability

[hongshi1]

Summary

This PR hardens Vertex streaming response parsing in ai-proxy.

Previously, Vertex streaming handling parsed each incoming chunk directly by splitting on newlines and unmarshalling line-by-line. That was fragile when an upstream SSE data: event was split across multiple network chunks. It could also drop the final payload because the last chunk returned [DONE] immediately.

Changes

• add buffered SSE data-line extraction helpers for streaming parsing
• make Vertex streaming response handling support split/incomplete chunks safely
• keep the final payload conversion on the last chunk, then append data: [DONE]
• add provider-level regression tests for:
  • split chunk reconstruction
  • final chunk payload preservation before [DONE]

Verification

go test ./provider -run 'TestVertexStreamingResponseBody|TestVertexStreamingResponseBodyBuffersSplitChunks|TestVertexStreamingResponseBodyKeepsFinalPayloadBeforeDone'
go test -run TestVertex ./...

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

gexuancheng seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}