fix: handle partial Redis rejection writes

[thweetkomputer]

What problem does this PR solve?

Issue Number: N/A

Problem Summary:

The Redis connection-limit rejection path uses non-blocking send() but assumes the complete error response is written in one call. A partial write can therefore truncate -ERR max number of clients reached before the accepted fd is closed.

This ports the follow-up fix from apache/brpc#3498.

What is changed and the side effects?

Changed:

• Continue sending the remaining plaintext Redis error after a partial write.
• Retry send() after EINTR.
• Stop immediately on EAGAIN, EWOULDBLOCK, or any other error, preserving the non-blocking accept path.
• Leave the pre-TLS rejection path unchanged.

Side effects:

• Performance effects(性能影响): The over-limit plaintext rejection path may perform more than one non-blocking send() after a partial write. Accepted connections, request handling, and TLS rejection are unaffected.

• Breaking backward compatibility(向后兼容性): None.

---

Check List:

• git diff --check origin/master...HEAD
• cmake --build /tmp/eloqdata-brpc-partial-build --target brpc_server_unittest --parallel 4
• Focused ServerTest coverage passed: dedicated-listener validation, plaintext rejection, pre-TLS rejection, and idle-connection slot recovery (4/4).
• The complete unit-test suite was not run.

The fresh unit-test configuration used a temporary CMAKE_PROJECT_INCLUDE compatibility macro because current master calls the removed use_cxx11() macro from test/CMakeLists.txt; that unrelated workaround is not part of this PR.