Improve linger loop

[essen]

erlang/otp#9529 (comment)

We should set {linger, {true, 0}} before the loop to ensure that the socket will get closed when the process ends (either at the end of the loop, or when killed).

[eagle-head]

Hi @essen, I'd like to take this one.

The fix itself per your OTP comment is mechanical — call setopts(Socket, [{linger, {true, 0}}]) after shutdown(write) in terminate_linger/1, mirrored in cowboy_http2.erl (same "copied from cowboy_http" block). What I'm unsure about is the test, since the suite is uniformly behavior-driven via CT.

Two shapes I considered:

1. Reproduce the leak from TCP Connection Ports Leaks erlang/otp#9529 end-to-end: handler that sends a large body, client reads with a small recv buffer so the server's output buffer stays unacked, kill the connection process, then assert the server-side port is no longer in erlang:ports() within a short window. Closest to the actual symptom users hit, but reproduction relies on kernel TCP buffering and may behave differently across Linux/BSD/macOS in CI.

2. Spin a listener with linger_timeout => infinity, drive it into the linger loop, and assert via inet:getopts/2 that the server-side socket has {linger, {true, 0}} set. Deterministic but asserts the mechanism rather than the observable behavior.

Is there a precedent in the suite I should follow, or a shape you'd prefer? Also: a single CT case in http_SUITE plus the mirrored fix in cowboy_http2:terminate_linger/1, or do you want a parallel test in http2_SUITE too?

Happy to go in whichever direction you confirm.

[essen]

I think we very much want number 1 even if the test doesn't work as well everywhere. If it shows that on Linux without the fix the socket doesn't get closed, and with the fix, it does, then that's good enough. We can leave the test for all platforms we just need it to demonstrate that it improves things.