Separate the ESMValCore internals, dealing with products, from
the core function operating on cubes.

This makes it easier to add a new preprocessor for ensemble statistics,
and also to make a new core function (lazily) compute statistics
across cubes.

This function operates directly on iris cubes. It supports
lazy evaluation, and requires less error-prone array handling.
However, it requires the input cubes to be very homogeneous, so
for backwards compatibility I kept the old function as well.

This way we can easily test and improve the new function until
we're happy, and then deprecate the old function.

Currently the tests fail - as expected. Reasons are that
'cell methods differ' or 'coordinates (time/air pressure) differ'.

We can discuss whether this should be addressed within or outside
of the multimodel/ensemble statistics funcions.

…gine