net: forward-port FreeBSD IPv6 stack with SLAAC and DHCPv6 autoconfiguration

[gburd]

This forward-ports the FreeBSD IPv6 stack to current OSv and adds the boot-time
autoconfiguration a unikernel needs (SLAAC and a DHCPv6 client), so an
IPv6-only or dual-stack application works out of the box. All IPv6 code is
compile-time gated so IPv6-less builds stay small.

What this does

• netinet6 stack — imports the FreeBSD 9 netinet6 sources (ip6, icmp6,
  nd6, in6_*, udp6, raw_ip6, mld6, frag6, dest6, route6, scope6) adapted for
  OSv. 45 new files under bsd/sys/netinet6/, plus the shared eth/tcp/udp code
  taught about IPv6.
• OSv fast paths for IPv6 — the net-channel classifier, LRO, and RCU
  connection lookup are extended to IPv6/TCP so v6 gets the same in-kernel fast
  path as v4 (mirrors the existing IPv4 optimizations).
• Linux socket compatibility — 43 IPV6_* socket options translated in the
  Linux compat layer (V6ONLY, PKTINFO, HOPLIMIT, TCLASS, MULTICAST_*,
  JOIN/LEAVE_GROUP, etc.), plus getifaddrs()/if_nameindex() over a NETLINK
  socket and AAAA-aware getaddrinfo.
• SLAAC — OSv has no rtsold, so the kernel now sends its own Router
  Solicitation (nd6_rs_output) at interface bring-up and processes the Router
  Advertisement to form a global address. Also fixes an ordering bug where
  ND6_IFF_ACCEPT_RTADV was never set on interfaces that attached before the
  accept-RA flag was enabled at boot.
• DHCPv6 client — a minimal RFC 8415 stateful client
  (bsd/porting/dhcp6.cc): SOLICIT/ADVERTISE/REQUEST/REPLY with DUID-LLT and
  IA_NA, plus Rapid Commit (two-message SOLICIT/REPLY). The loader chooses SLAAC
  vs DHCPv6 from the RA flags: an Autonomous-prefix RA (radvd, QEMU SLIRP) uses
  SLAAC, a Managed-flag RA (many enterprise/cloud networks) uses DHCPv6. Works
  with standard servers (isc-dhcp, kea, dnsmasq).
• Boot options — --ipv6 selects the IPv6-only boot path (skips the
  blocking DHCPv4 wait); static --ip/--defaultgw/--nameserver already
  accept v6 addresses.

Compile-time gating

Everything is behind conf-INET6 (the FreeBSD INET6 macro) and, for the
DHCPv6 client, conf_networking_dhcp6. A conf-INET6=0 build compiles and
links with no IPv6 code and no dangling references (verified), saving ~2.7 MB.
The default remains conf-INET6=1.

Testing

• tst-tcp-v6-smoke (new, no boost): IPv6 + IPv4 loopback TCP echo, concurrent
  dual-stack, and the socket-option/multicast/getaddrinfo API surface. Passes.
• tst-ifaddr-dump (new): dumps interface addresses to observe SLAAC/DHCPv6.
• tst-ipv6-internet (new, manual): DNS AAAA + IPv6 TCP + HTTP to a real host.
• Existing tst-tcp-v6, tst-pktinfo, tst-socket-timestamp build and run.

Validated on KVM/QEMU (SLIRP ipv6=on → SLAAC) and on a bare-metal host:

• SLAAC forms a global address from a Router Advertisement.
• The DHCPv6 client obtains a global address from a standard dnsmasq DHCPv6
  server (SOLICIT with Rapid Commit → REPLY → address bound), confirming the
  client is standards-compliant and not tied to any specific network.
• IPv6-only boot (no IPv4 on the link) reaches the IPv6 internet.

Notes for review

• The history preserves the original ipv6-branch commits (FreeBSD import →
  OSv adaptation → net-channel → Linux compat → getifaddrs/netlink → tests) so
  authorship is intact, with the SLAAC + DHCPv6 work as the final commit. Happy
  to squash if preferred.
• The DHCPv6 client lives in bsd/porting/ (not core/) because it needs the
  BSD-internal netinet6 headers, which cannot coexist with the musl/glibc
  in6_addr that boost::asio pulls in; the DNS setter is reached through a
  small C-linkage shim in libc/network/__dns.cc.

[gburd]

Follow-up (deliberately not in this PR): IPv6 inherits OSv's high-value fast paths — net channels (classify_ipv6_tcp, tcp_net_channel_ipv6_packet), LRO (ETHERTYPE_IPV6 in tcp_lro.cc), the RCU connection classifier, and the AF-agnostic TCP header-prediction path. What it does not yet inherit is virtio-net hardware checksum + TSO offload for IPv6: drivers/virtio-net.cc hard-codes the offload switch to ETHERTYPE_IP (TX builds only GSO_TCPV4, RX has a literal // How come - no support for IPv6?!), and if_hwassist/if_capabilities advertise IFCAP_TSO4 but not IFCAP_TSO6 or the CSUM_*_IPV6 bits.

Net effect: IPv6 is correct but computes checksums in software and doesn't get large-segment TSO, so high-throughput v6 sends/receives use more CPU than v4. This is a self-contained, driver-local performance follow-up (extend the TX/RX offload switch to ETHERTYPE_IPV6 → CSUM_TCP_IPV6/GSO_TCPV6, add IFCAP_TSO6), best done and benchmarked as its own PR rather than expanding this correctness-focused one.

[gburd]

Heads-up before merging: this branch was last synced against an older master (via a merge commit), so although GitHub reports it mergeable, its tree predates several tests that have since landed (tst-iovcnt-guard, tst-splice, tst-fs-syscalls, tst-membarrier, tst-mremap) and it still edits the tests := literal in modules/tests/Makefile directly. A clean rebase onto current master hits genuine net-stack conflicts (net_channel.cc, tcp_input.cc, routecache.hh) from the merged IPv6 net-channel/classifier changes, so this needs a real forward-port rather than a mechanical Makefile fixup. I will rebase it onto current master and convert the test entries to standalone tests += lines (per #1469) before it is ready; please hold this one until then. The smaller syscall/mm/ext4 PRs are already converted and ready.