quic-go will tear down connections if sendmsg fails with ENETUNREACH

I'm not sure if this should be fixed but we did run into this on systems where dhcp was configured to recreate the default route for every rebind.

We could fix it by tweaking send_conn.go's Write to treat ENETUNREACH and UHOSTUNREACH as transient errors by just returning nil. The packet would be resent and if the network was truly down eventually the probe failures would tear the connection down.

I don't believe TCP connections are torn down by the default route disappearing for a tick like that. However the simpler fix was just to tweak our dhcp script to keep the route if the gateway was unchanged during a rebind.