Fix SCTPListener.Close race with concurrent Accept

[igor-lazic]

Summary

SCTPListener.Close currently calls syscall.Shutdown and syscall.Close on ln.fd without any synchronisation. Concurrent AcceptSCTP and Addr calls read ln.fd directly. This produces three observable bugs that are easy to hit in any program that calls Close while a Serve loop is blocked in Accept:

1. A goroutine blocked in accept4() returns EINVAL or EBADF when Close runs concurrently, instead of a sentinel callers can detect (cf. net.TCPListener, which has returned net.ErrClosed since Go 1.16). Real-world failure: downstream callers like go-diameter log every graceful shutdown as "accept error: invalid argument" or "accept error: bad file descriptor".
2. AcceptSCTP after Close calls accept4 on the now-dead fd integer. If the OS has reassigned that fd between Close and Accept, accept4 either fails or, worse, returns an accepted fd from an unrelated socket.
3. AcceptSCTP always returns a non-nil *SCTPConn (wrapping fd=-1 on the error path), violating the (net.Conn, error) API contract. Callers that defensively invoke methods on the returned conn hit another EBADF from setsockopt.

The fix

The change mirrors the pattern this package already uses for SCTPConn — same field shape, same atomic primitive, same close protocol:

Element	SCTPConn (existing)	SCTPListener (this PR)
Field	_fd int32 (sctp.go)	_fd int32 (sctp.go)
Accessor	c.fd() via atomic.LoadInt32(&c._fd)	ln.fd() via atomic.LoadInt32(&ln._fd)
Close	atomic.SwapInt32(&c._fd, -1) (sctp_linux.go:157)	atomic.SwapInt32(&ln._fd, -1)

Specifically:

• Rename SCTPListener.fd → _fd int32, accessed via atomic.LoadInt32 / atomic.SwapInt32.
• Close uses atomic.SwapInt32(-1) so only one caller observes the real fd; subsequent calls return syscall.EBADF without re-issuing syscall.Close on a possibly-reused fd number.
• AcceptSCTP and Addr take an atomic snapshot of _fd and bail to ErrListenerClosed / nil if the listener has been closed. When accept4 was already blocked inside the kernel when Close ran, the post-call atomic re-check normalises the resulting EBADF / EINVAL to ErrListenerClosed.
• AcceptSCTP now returns (nil, err) on every error path.

ErrListenerClosed is a new exported sentinel, defined with errors.New to stay compatible with the existing go 1.12 directive (no net.ErrClosed / errors.Is usage). Callers compare with ==.

Test plan

• New sctp_listener_close_test.go: Close-during-Accept, Accept-after-Close, double-Close, 16-way concurrent Close, Addr-after-Close
• Verified RED on master: the first two tests fail with exactly invalid argument (EINVAL) and bad file descriptor (EBADF), with a non-nil *SCTPConn{_fd:-1} returned alongside the error — i.e., the test reproduces the downstream go-diameter failure mode byte-for-byte
• All five tests pass with this patch
• Existing test suite remains green under go test -race -count=1

What this PR does not touch

• sctp_unsupported.go (the non-Linux stub still returns ErrUnsupported)
• go.mod (still go 1.12)
• SCTPConn.Close's pre-existing if fd > 0 check (mildly off — should be >= 0 since stdin is a valid fd — but unrelated to this race fix, happy to send as a separate PR)