Based on the kernel version, can be overridden explicitly.

Prefixes are counted in bits, not in bytes.  It is safe to compare all
16 bytes, since prefixes are always normalised -- they have zeroes beyond
the prefix length, see mask_prefix.  Alternatively, we could use
prefix_cmp.

Was used in the first version of the disambiguation.c algorithm.

This reverts commit 9edf0a14.

Thanks to Pierre Pfister.

When restarting network on openwrt, the rules were flushed.  In
this patch, we ask the kernel for rule changes, and try to
reinstall (or flush) the rules, as it should be.

Since the kernel behavior is the one desired for ipv6 (match by destination
first, then source), we don't install disambiguation route entries for ipv6.
The behavior for v4 doesn't change.

The linux kernel should be able to deal correctly with source-sensitive ipv6
route entries since 3.11 (when it is configured with IPV6_SUBTREES).  This patch
modify the kernel so that ipv6 route entries sent into the kernel are not pushed
in a different routing table.

The behavior for v4 doesn't change.