fix: handle internal API stream chunk boundaries

[snowyukitty]

Description

Fixes #2177.

The internal inference client previously decoded and parsed each transport
chunk independently. A chunk can end in the middle of either an SSE data line
or a multi-byte UTF-8 character, so a valid streamed response could fail with a
SyntaxError and discard the rest of the answer.

The parser now keeps the TextDecoder in streaming mode, buffers text between
reads, and parses only complete newline-delimited frames. It also flushes the
last unterminated frame at EOF and skips an isolated malformed data line while
preserving the existing error-frame behavior.

Regression tests feed the response one byte at a time through a multi-byte
character and verify that a malformed line does not prevent later valid
content from streaming.

Type of Change

• Bug fix
• New feature
• Documentation
• Other (refactor, build, chore)

How to test

1. Run npm ci.
2. Run npm test.
3. Run npm run lint.

The focused regression suite is:

npx vitest run client/modules/textGenerationWithInternalApi.test.ts

Checklist

• npm run lint passes
• Tests pass (npm run test), with tests added where it made sense

[felladrin]

Excellent! Thank you for this!