vio: Increase rx mbuf size with lro

vio: Increase rx mbuf size with lro

bluhm found that using bigger rx mbufs helps tcp splice performance if
lro is enabled. Use 4k in that case.

Also fix confusion in rx dmamap segment count. Even with lro/tso, we
only put unfragmented mbufs into the rx queue. Therefore we only need
max. 2 segments, one for the mbuf and one for the separate header for
legacy virtio devices.

OK bluhm@