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`

This is required for the bindings, because the 2D views on 4D tensors
must be on the correct type (i.e. pytensor/rtensor/xtensor). Although,
casting of the views to become xtensor could have been used instead.

In doing so, there is no source file to be compiled anymore, so this
becomes a header-only library. But to keep the "public" headers separate
from the implementation, a subdirectory "detail/" is added.

In order to harmonise with `evalp`, and in view to use `evald` in `evalp`,
the subsets dimension (corresponding to the temporal masks) needed to be
explicitly included in the output of `evald` rather than allowing a
limited number of masks corresponding to the number of time series
evaluated to be used.

The bootstrapping method is based on a non-overlapping block sampling
with replacement, where the blocks are years of data. The number of
samples and the sample length (i.e the number of year blocks) are both
customisable.

The method is accessible both for deterministic and probabilistic
evaluation where a new axis is added. For now, the metrics for all the
samples are returned, but in the future, some summary statistics would
be implemented (e.g. quantiles or mean/standard deviation).

/!\ For determinist evaluation, the n-dimensional functionality became
    untenable such that the number of dimensions was fixed and
    restricted to 2D tensors.

New unit tests are included to test both the bootstrapping generator
and the numerical results obtained with the bootstrapping turned on.

The use of `xexpression` and template parameter was showing its limits
as the data type could not be different between the different n-dim
inputs (e.g. masks had to be as double rather than bool). Since there is
no real need to be able to use `evald` on `xarray`, the genericity of
its API is reduced so that it now expects inputs to be tensor only. The
number of dimensions for the tensors becomes a template parameter.

both for deterministic and probabilistic input data

Unlike 3de89046, this also deals with
missing predictions. This effectively pairwise remove those time steps
where either observations or predictions are missing data. This does so
without creating copies of the input data, via the use of xfunctions
stored as class members.

The unit test for missing data is updated to include NAN also in the
predictions (with a varying number of NAN across the 5 time series).

In order to avoid copies of the input data, xfunctions are generated
to replace timesteps in predictions where they are NaN in observations.
This means that with this implementation, missing predictions are not
supported.

See evalhyd-r#1
for details on rtensor bug. Once this bug is fixed, `xt::stddev` should
be brought back.

With xexpression, the compiler will complain if the type used does
not implement the xexpression interface. Before the move to OO approach,
it was present, but got lost along the way.

It seems that good practice for C++ applications is to only include
*public* headers in "./include" folder and keep source files and
*private* headers and implementation source files in "./src".

The package now relies on Evaluator objects to store and compute
elements and metrics. This makes it easier to avoid unintended copies.