This is an attempt to add access to determinist metrics to probabilist
evaluation (i.e. via the choice of a reducer on ensemble members).

The idea is to instantiate one determinist::Evaluator per site (in order
to reduce the dimension to 2D) as part of probabilist::Evaluator. This
way, the determinist evaluator handles all the memoisation/dependency
management, while the probabilist evaluator only needs to get the metrics
via a call to the getter methods of the determinist evaluator (e.g.
`get_RMSE`).

Unfortunately, while this commit compiles/links, it fails to run because
the views to q_obs/q_prd given to each determinist evaluator lose the
reference to the data as soon as it leaves the constructor method. So
there is likely a move/forward required somewhere, or the use of
a constructor method expecting universal references, but for now I could
not find a working solution along those lines. So I park this here for
now.

In view to be able to use determinist metrics in `evalp`, the temporal
masks need to feature a dimension specific to each prediction series.
In fact, this was already done internally to `determinist::Evaluator`
to deal with potentially missing predictions. From a determinist point
of view, this is actually beneficial as well since (a) it allows the
user to use different masks for different series and (b) it avoids an
internal copy of masks to go from 2D to 3D.

only showing up when multiple lead times provided

Until now, the probabilist Evaluator was not able to compute multi-sites
and/or multi-leadtimes metrics, because it was given one site and one
lead time at a time. This was done to spare some memory, but it was not
logical to load all sites and all lead times for the input data and then
only process a small chunk at a time.

There is no multi-sites/multi-leadtimes metrics implemented yet, but
the Evaluator is ready for it now. This implementation makes full use
of the broadcasting power of xtensor.

while Brier-based scores are symmetric, Contigency Table-based metrics
are not, so that a definition how what an "event" means is required.
A new optional parameter *events* is added to `evalp` (taking either
"high" or "low" as value).

as mentioned in https://github.com/xtensor-stack/xtensor/issues/2629,
the behaviour of the random generator is controlled by the implementation
of the standard library, so it cannot be expected to behave the same
across platforms

since those default values were only used to check whether these
parameters were provided or not, they are turned into optional
parameters

so that specifying the template arguments become optional

GNU GCC was complaining about a missing type definition ("no type name
'type' in 'struct::enable_if<false, int>'") for the false case
(as explained here https://stackoverflow.com/a/12625334). So given that
there are two/three false conditions for `evald`/`evalp`, it seemed
better to drop these rank checks from the templates, and use the
function `xt::get_rank` instead.

including:
- extracting the element/intermediate/metric calculations from the
  `determinist::Evaluator` class
- introducing new getter methods in `determinist::Evaluator` class
  that handle the element/intermediate/metric inter-dependencies
  internally to the object
- removing external inter-dependencies handling from `evald`

including:
- extracting the element/intermediate/metric calculations from the
  `probabilist::Evaluator` class
- introducing new getter methods in `probabilist::Evaluator` class
  that handle the element/intermediate/metric inter-dependencies
  internally to the object
- removing external inter-dependencies handling from `evalp`