Refactor IPv6 generator to support hosts with smaller than prefix 64

[furkansahin]

Some providers place multiple hosts behind the same /64. Therefore, the
gateway is the same for both servers. We then try to use this whole /64
in Ubicloud and carve out /80 addresses to assign to VMs. However, when
a single server claims the full /64, we cannot add the second host. Here
in this PR we solve this problem by supporting subnets that are smaller
than /64. This way, when there are 2 servers behind the same subnet, we
can selectively assign a (let's say) /68 subnet to one and another /68
to another. After that, we can still carve out /80 addresses from those.

One important change we are making is in the ip6_random_vm_network. That
is needed because, the rounding up causes the fail line to be
triggered. Here is a scenario;

• The host_prefix is 68.
• The prefix is 79
  With the rounding up, we produce a lower_bit that is 2 bytes. These 2
  bytes before the /79, doesn't fit together with /68 prefix, we may reach
  an address that is not directly a child address of the host_prefix.
  Therefore, we keep that at minimum 1 and produce a number that is a
  single byte which easily fits under the /68 prefix.

[fdr]

I'm a bit puzzled by this patch. It appears it changes how many random bits are used in all cases, e.g. a reduction in precision from n to C that is unconditional. Is that what we want?

Also, is bytes_needed = [((subnet_bits - 1) / 8), 1].max fixing a bug?

also, does this patch more or less do everything necessary to support longer prefixes?

[furkansahin]

So, when we go to the longer prefixes landscape, we're talking about > 64 and < 80. If we have the example of 68,

bytes_needed = ((subnet_bits - 1) / 8) + 1

would return 2 bytes which is longer than we needed, and in return we hit the line for most of the proposals;

fail "BUG: host should be supernet of randomized subnet" unless net6.rel(proposal) == 1

By adding [].max, we also make sure we won't get nil from SecureRandom.bytes(bytes_needed).unpack1("C") for cases bytes_needed=0.

So, without the changes in this PR, we would either hit nil ref in line 84 or hit the fail in line 92 for hosts with longer prefixes.

[furkansahin]

@fdr waiting for your input here

[fdr]

I would rewrite the commit message, but I'm out of time. I think it is somehow repetitive and could be condensed.

And, it omits that ALL random subnets, even for hosts with shorter prefixes, will use fewer random bits as a side effect.

Secondarily, that you are fixing a bug, as the previous code would crash if host_prefix was too large, rounding the number of random bytes generated to zero.

But it's better to commit in a timely manner than not, so, here's the approval bit.

[fdr]

Actually, now that I think of it, why even bother with the bytes needed with the "C" unpack? Can you avoid all that and generate, always, one byte of random bits? So the code can stop pretending it handles more or less random bytes?

[furkansahin]

I have updated the commit message, thanks!