overlay: replace per-fd tun readers with a batched Queue interface

Device loses io.ReadWriteCloser + NewMultiQueueReader in favor of
Queues(n), which returns up to n tio.Queue objects; platforms without
multiqueue hand back their single queue and the interface sizes its
reader routines to what it actually got. Queue.Read returns a batch of
borrowed packets (single-element for every current backend) so a future
backend can deliver more than one packet per syscall without another
interface change.

The Linux poll/eventfd machinery moves out of tun_linux.go into the new
overlay/tio package: nonblocking fds, a shared shutdown eventfd owned by
the queue set, and pollfd arrays built on the stack so concurrent
writers parked in blockOnWrite no longer share Revents storage. Other
platforms wrap their existing one-datagram Read/Write in a singleQueue
adapter that owns a private scratch buffer, so multiqueue-by-sharing
devices (user, disabled) no longer race concurrent readers on one
buffer.

This is the tun-interface subset of better-tun-interface-ordering,
extracted at 18dc13b with none of the GSO/GRO offload mechanics and no
udp/sendmmsg changes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

inside.go

@@ -37,7 +37,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			_, err := f.readers[q].Write(packet)
+			_, err := f.queues[q].Write(packet)
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -96,7 +96,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.readers[q].Write(out)
+	_, err := f.queues[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}