This reverts commit b11c4610.

when observations only contain NaNs, the resulting data subset is
empty, which makes the quantile calculation fail

this can happen when evalhyd is used on a period with no observations
(which may rest on the user responsibility), but this can also happen
when subperiods (i.e. via temporal masking) have no observations even
if there are observations in the overall period (which is less easy
for the user to anticipate)

Low flow events were defined as the complement of high flow events,
meaning that Brier scores were symmetric. But it is more consistent to
include the threshold value in both the definitions of low flow and
high flow events since we either study one or the other, and rarely
both at the same time, especially not using the same threshold values.
Plus, the choice of including the threshold in high flow events rather
than low flow ones was really arbitrary.

Note, only the unit tests for contingency table-based metrics are
impacted because threshold-based metrics are using evants="high".

in particular, BSS was returning 0 when BS reference was null, but 0 is
a meaningful value for such score, so it is preferable to return -INF
to indicate that the denominator was null and hence the division
undefined

the same philosophy is also used where relevant (i.e. in AWN, AWI, WSS)

in `xt::stddev``, when passing a scalar, it is passed to *ddof*
instead of *axes*, so to compute the standard deviation on a given axis,
the axis must be passed in curly braces

it is also a dependency of xtensor, but evalhyd uses it directly too

related to e453205d

because of the random process in the rank computations, and the
different input tensor sizes

replace clumsy xt::sum ... == xt::scalar, by xt::all