fix: make SCTPWrite non-blocking to avoid stalling on a full send queue

[gomaja]

Split out of #87, where it was buried in an unrelated close-logic change. It alters SCTPWrite's contract for every caller, so it is better reviewed on its own.

SCTPWrite called sendmsg with no flags, so it blocked whenever the send queue was full. A peer that stops reading, or goes away without tearing the association down, parked the writing goroutine indefinitely and left the connection in CLOSE_WAIT with no way for the caller to recover. Passing MSG_DONTWAIT returns EAGAIN instead, so the caller decides what to do.

Compatibility

This is a behaviour change worth stating plainly: callers that relied on the write blocking until there was room must now handle EAGAIN and retry. A plain write loop that ignores the error will drop messages.

EAGAIN is also not limited to a peer that has stopped reading — a writer that outruns an actively reading peer sees it too. The tests measure this rather than assert it away:

• with a peer that has stopped reading, the writer reports EAGAIN instead of hanging
• against an actively reading peer, roughly one write in twenty returns EAGAIN once the writer gets ahead
• a retry loop delivers all 200 messages

The middle case is the one existing callers will notice, so it is a test rather than a comment.