resolves #10

fixes #5

fixes #9

resolves #8

resolves #7

related to evalhyd-r#6

for arrays, rank is not applicable, so it should not be checked

it had no effect on numerical results

`xt::median` and `xt::quantile` are not NaN-proof, and do not have
equivalent NaN functions (like `xt::mean` and `xt::nanmean`) so
it required a filter to drop the NaN before computing the corresponding
condition value

e.g. "q_prd_median{>quantile0.3,<quantile0.7}" was not working
because the string was too long, but the regex were also quite
complicated, so this revises the approach and simplifies the regex

but to overcome the length limitation with C++ regex, "quantile" now
need to be specified as "qtl" to be shorter

the energy score (ES) is a multi-variate (multi-site)
generalisation of the CRPS

new metric is called CRPS_FROM_ECDF

new metric is called CRPS_FROM_BS, and existing metric is renamed
CRPS_FROM_QS

available in xtensor 0.24., see https://github.com/xtensor-stack/xtensor/issues/2651

Checking the "nan-ness" of an integer is ill-posed, such that
the standard library of some compilers returns an error message.
Dropping the nan filter here for the sum is safe because `xt::equal`
evaluates to false when either operand is NAN.

where data completeness corresponds to the actual number of time steps
available for each subset/sample, once missing data, mask, and bootstrap
have been applied

The template argument defining the type of for temporal masks was not
used in the elements/intermediates/metrics functions, and since a copy
from the user-provided input masks is required to not modify it, it
makes more sense to directly initialise it as an xtensor rather than
propagate a template argument to use pytensor/rtensor.

by extracting the relevant variables from existing BS_CRD

because `xt::sort` does not behave nicely with series containing NaN
values (e.g. it does not consistently put NaN values at the end/
beginning), so need to eliminate them before the sorting, which requires
to treat each series separately

by storing errors before computing quadratic errors