see evalhyd-cpp#7

and add unit tests to cover these cases (which was not the case before),
also add a unit test for special case with no subset (i.e. "t{:}")

existing CRPS is also renamed CRPS_FROM_QS

this way, all 51 members can be used without clogging the scripts

which required to update to the latest evalhyd commit and adjust the
expected output for contingency table-based metrics

to work with different site/leadtime series and thus making sure that
there is no silent broadcasting resulting in the correct metric value
even if it worked with the wrong site/leadtime

to make sure that there are no silent broadcasting errors because
of size-one dimensions

otherwise it is not very easy from the error message to deduce the
problem is due to a mismatch in number of dimensions

which includes a temporal masking functionality for deterministic
evaluation

which includes handling of `numpy.nan` for missing data in observations
and predictions. This is documented in the `evalp` docstring and new
unit tests are added to check this new functionality.

which includes the explicit assignment of NaN in outputs where threshold
is not provided by user for given site (i.e. set to NaN)

version which adds "sites" axis to dimensions of thresholds and masks