Fix GH-24229: fcgi_close() passes shutdown() a literal 1

[mpyw]

Fixes GH-24229.

The quoted description below, and the checks it describes, were done with an LLM (Claude). I reviewed them.

fcgi_close() half-closed the connection with shutdown(req->fd, 1), meaning SHUT_WR, as the comment after it says. Every platform PHP supports uses 1 for SHUT_WR, but POSIX does not fix the value. WASI uses 2, and 1 is SHUT_RD there.

With the read side shut, the drain loop ended at once. close() could then run with the web server's last FCGI_STDIN unread, and Linux reset the connection. That discarded the response before the web server read it.

This uses SHUT_WR, which <sys/socket.h> already provides there. The _WIN32 branch is left as is: 1 is SD_SEND, which is right on Windows.

There is no test: on the supported platforms, SHUT_WR is 1, so nothing changes. I checked that in a build of this branch, configured with --disable-all --enable-cli --enable-cgi --enable-fpm on Debian trixie. sapi/cgi/tests and sapi/fpm/tests give the same results with and without this change: 132 passed, 26 skipped, and the same 2 failures, which come from running them as root in a container.

The fix itself was checked with php-cgi built for wasm32-wasip1, in FastCGI mode on wazero. On Linux, 10 of 600 requests failed before it, and none after.

This targets PHP-8.4 as the lowest branch with bug fixes. I am happy to retarget it to master if you see this as a portability change rather than a bug fix.

[ndossche]

You're correct, and SHUT_WR==1 on Linux, thanks

[LamentXU123]

Thank you!