We did not yet set a defined level before testing loggers. Therefore
results were unpredictable.

We will check at least whether the man page is current and ensure that
at least words from help output are mentioned in man page.

When using `realdice` randomness source and picking values out of short
sequences (sequences shorter than the number of dice sides), we might
require additional dice throws (if a value rolled is not in range of
sequence length). We now print a hint when this happens.

We avoid the modulo operation now, if it comes to small sequences.
Instead we discard invalud rolls and require new ones.

In order to keep all entropy we could in case of sequence lengths
unequal to powers of the number of dice sides just discard rolls and
require further ones, until we get a number in the allowed range. But
this could potentially require an endless number of rolls.

Instead we implement the following policy:
If R=log(<sequence_len>) with <number_of_dice_sides> as basis is >= 1,
we require int(R) rolls. We will ignore elements in the sequence with
index > int(R).

Sample: For 6 dice sides and a sequence of 215 (= 6**3 - 1) elements R
would be log_6(215) = 2.9975... but we would pick an element from the
range [0..35], i.e. (6**2 - 1).

If R is 0, we will require exactly one roll, if <sequence_length> is a
divisor of <dice_sides>.

In all other cases, <sequence_length> is not a divisor of <dice_sides>
and <sequence_length> is shorter than the number of dice_sides. In this
case we will roll dice until we get a result in [1..<sequence_length>].
Other rolls will simply be discarded.

Sequences with lenths != s**n (with s:= number of dice sides, n := any
natural number) require an extra roll.

The number of rolls is cruical for the whole `choice` distribution and
a specialized function is much better testable.

Wordlists that contain dice_sides**n elements will always be distributed
equally by `choice` when using real dice.

of dice-based random.

And place the new, integrated fixture on top. We keep it in this
module (instead of moving it to conftest.py) because it contains
functionality related to randomsources only.

Remove usage of old classmethod and the method itself.

This one is supposed to replace the existing classmethod and should become a regular fixture.

Looks like there are no problems with Python 3.6 yet. We make it a
regular target in tox and will start official support in `setup.py`
after a while.

Tell about the new default wordlist.

The EFF list is currently certainly the scientifically most advanced
list. Thanks to @anarcat for proposing this switch! The securedrop
list is still available.