This is based on Henry's initial patch from GH-166. Changes:

* Fixed the return type of cidr_merge() so it's always IPNetworks,
  not IPRanges as it was in some cases in the original patch
* Added tests
* Updated documentation

This addresses GH-171.

Co-authored-by: Henry Stern <hstern@securityscorecard.io>

This is basically PR https://github.com/netaddr/netaddr/pull/151 with
added tests.

Co-authored-by: xpac1985

IPv4 has those tests already so adding for coverage parity.

This reduces the amount of work in the loop which simplifies the loop
logic and also has a performance benefit. Benchmarked with:

    % PYTHONPATH=. python -m timeit -n 5 \
        -s 'from netaddr import IPNetwork, spanning_cidr' \
        -s 'import random' \
        -s 'hosts = list(IPNetwork("192.168.0.1/12"))' \
        -s 'random.shuffle(hosts)' \
        'spanning_cidr(hosts)'

Before:

    5 loops, best of 5: 4.95 sec per loop

After:

    5 loops, best of 5: 4.6 sec per loop

(Tested with CPython 3.8.3 on mid-2015 MacBook Pro with 2.8GHz Intel Core i7)

Closes #205

From Jakub: this is a performance improvement. A simple benchmark
demonstrates this:

    % PYTHONPATH=. python -m timeit -n 5 \
        -s 'from netaddr import IPNetwork, spanning_cidr' \
        -s 'import random' \
        -s 'hosts = list(IPNetwork("192.168.0.1/12"))' \
        -s 'random.shuffle(hosts)' \
        'spanning_cidr(hosts)'

Before:

    5 loops, best of 5: 35.4 sec per loop

After:

    5 loops, best of 5: 4.95 sec per loop

(Tested with CPython 3.8.3 on mid-2015 MacBook Pro with 2.8GHz Intel Core i7)

Like RFC 3021, IPv6 defines point-to-point subnets that must
be handled separately, i.e. don't reserve first IP address.

This patch aims to implement this, while refactoring code in
iter_host function to reduce code duplication.

Tests for this feature also added to ensure there is no regression.

Signed-off-by: Damien Claisse <d.claisse@criteo.com>