[STACKED on #1442] virtio-net: enable IPv6 checksum and TSO offload

[gburd]

Note

Stacked on #1442 (the IPv6 forward-port). This branch is based on
pr/ipv6-forward-port, so the diff below is cumulative and re-includes the
IPv6 stack. Once #1442 merges I will rebase onto master and this reduces to
the two-file drivers/virtio-net.{cc,hh} change. Please review #1442 first.

Performance follow-up to #1442. IPv6 already inherits OSv's high-value fast
paths (net channels, LRO, RCU connection classifier, AF-agnostic TCP header
prediction), but the virtio-net hardware offload paths were hardcoded to
ETHERTYPE_IP:

• TX built only VIRTIO_NET_HDR_GSO_TCPV4;
• RX rejected everything but IPv4 (with a literal // How come - no support for IPv6?! comment);
• if_hwassist/if_capabilities advertised IFCAP_TSO4 but not IFCAP_TSO6
  or the CSUM_*_IPV6 bits.

So IPv6 computed checksums in software and never used TSO, burning more CPU than
IPv4 on high-throughput sends/receives even though tcp_output already emits
CSUM_TCP_IPV6.

This change:

• negotiates VIRTIO_NET_F_HOST_TSO6 / VIRTIO_NET_F_GUEST_TSO6;
• advertises IFCAP_TSO6 and CSUM_TCP_IPV6 | CSUM_UDP_IPV6;
• TX: adds an ETHERTYPE_IPV6 case (csum_start past the 40-byte base header,
  gso_type = VIRTIO_NET_HDR_GSO_TCPV6); packets with extension headers have
  ip6_nxt != TCP/UDP and fall back to software checksums;
• RX: accepts ETHERTYPE_IPV6 in the checksum-offload path (the TCP/UDP
  checksum-offset logic is identical for v4 and v6);
• uses large receive buffers when GUEST_TSO6 is negotiated.

Boot-tested with IPv6 SLAAC over virtio-net (loopback + concurrent dual-stack
TCP pass, global address configured through the modified TX/RX paths). A
throughput A/B (v6 with vs without offload) is the natural benchmark once the
base merges.