tnet: keep udp poller active after send errors

[Rememorio]

Summary

• keep UDP listener fds registered with the poller when a queued datagram fails with a non-retryable sendmmsg error
• drop the failed UDP datagram and continue flushing later queued datagrams
• add a Linux regression test covering a failed queued UDP packet followed by a valid packet

Root Cause

UDP async writes can enqueue packets in outBuffer and rely on poller write events to flush them. If sendmmsg returns a non-EAGAIN error for the queued head packet, udpOnWrite previously returned that error to the poller. The poller then detached the shared UDP listener fd, while the UDP connection object stayed active with writing held and queued packets left in outBuffer.

Validation

• go test -run TestUDPConnWriteToDropsFailedPacketAndContinues -v -count=1
• go test -run 'TestUDPConnWriteToDropsFailedPacketAndContinues|Test_netFD' -v -count=1
• go test ./...

Fixes #26

[WineChord]

LGTM