While the backend stores the information as floats, the calculations
don't have to do be affected by this. For example:

0.3 + 0.3 + 0.3 == 0.8999999999

Using decimals improves the calculation, and avoid falling into the
wrong weight band.

This also removes the attr() marking of tests. This can be implemented
in pytest with marking the tests with a custom marker if required.

Making it overridable allows altering the weighing method of the basket,
which I wanted to do in a project. As get_classes doesn't support
importing from top-level products, it had to be moved to a scales
module.

Other class names tend to use singular, so as the imports had to be
changed anyway, Scales was renamed to Scale.

We no longer need to pass the StockInfo instance as the basket has a
reference to the strategy class and can fetch it for itself.  This means
we don't have to break backwards compatibility with Oscar < 0.6.

Split into unit, integration etc.