- Improve the docstring in arCalcHierarchical
 - Remove a redundant line in arInitHierarchical

Previously, this argument was accepet but unused by arCalcHierarchical.
Now it is used to avoid computing scale gradients and sensitivities of
observables when not necessary.

Previously, arInitHierarchical was using isSymType, the latter being a
function introduced in R2019a. This commit removes this dependence and
instead implements a small custom function which does the trick.

The scale parameters suitable for hierarchical optimization are those
which appear _only_ in observables of form `s*xz` where `s` is the
subject scale parameter and `xz` is a model species or a derived
variable. Until now, arInitHierarchical has been detecting parameters
which appear _at least in one_ observable of form `s*xz` (meaning that
they were allowed to appear also in obervables of other form).
Consequently, parameters not suitable for the hierarchical optimization
could be detected by arInitHierarchical and the results of the
hierarchical optimization could be not correct. This commit fixes it by
adding extra checks in arInitHierarchical to make the latter function
detect only the right parameters.

The implementation of hierarchical optimization depends on the specific
form of the merit function. We currently assume the following one:

    J(s,p) = \sum_j (data_j - s*x(p,t_j))^2/std_j^2

Each feature that modifies this expression also breaks the method of
hierarchical optimization, making its results incorrect. To avoid this,
in arCalcHierarchical we try to check if d2d features affecting the form
of the merit function are enabled. This commit adds extra checks of this
kind in arCalcHierarchical, hopefully making the checklist complete (or
close to complete).

The removed settings changes are redundant since these settings are
tested anyway in arCalcHierarchical.

Currently arCalcHierarchical requires no error fitting, meaning that
there can be no dependence of errors on any fittable parameters. Until
now, arCalcHierarchical has detected error fitting incorrectly,
excluding some possible settings which actually had no error fitting. In
other words, the test for error fitting has been to strict. This commit
fixes the issue by implementing a better test for error fitting in
arCalcHierarchical.

This commit updates the help text of arInitHierarchical to reflect the
changes in its functionality introduced in commit c3bbe5a. This should
had been done in that commit. Additionally, the present commit fixes a
typo in the subject help text.

Hierarchical optimization is not feasible if the solution of the model
is zero for all measurement times. This commit adds the respective
assertion in arCalchierarchical.

This extra assertion is not strictly necessary. Degeneration of the
solution to zero would result in NaN scales, subsequently resulting in
NaN residuals, which would trigger an error anyway. Nevertheless, with
the extra assertion we get a more specific error message which may make
it easier to debug.

Currently arCalcHierarchical requires no dependence of errors on any
fittable parameters, neither explicitly nor implicitly. Until now,
arCalcHierarchical has been asserting that there is no fitting of error
parameters, but at the same time has been allowing implicit dependence
of errors on fittable parameters of the model via observables. This is
a bug since with such implicit dependence of errors on fittable
parameters the results of arCalcHierarchical are incorrect. This commit
fixes it by asserting the use of experimental errors, which are
independent of the fittable parameters by definition.

At the moment, arCalcHierarchical allows experimental errors only, so
the lines which handle the case of model-based errors are redundant and
obscure the code. This commit removes such lines.

See also commit 1b7fb04.

In arInitHierarchical, the cell arrays xSyms and zSyms has been
created assuming that there is only one model loaded (i.e. the length of
the ar.model structure is 1). This commit corrects the support for
multiple models in arInitHierarchical.

For certain purposes, in arCalcHierarchical we have been testing if all
elements ar.model(im).data(id).useHierarchical are NaNs. But this cannot
happen since the array useHierarchical is constructed such that its
elements may be either false or true (and not NaN). So we should rather
test if the elements of useHierarchical are all false. This commit fixes
this issue.

Until now, arCalcHierarchical has been working correctly only for data
series with no duplicate time points (or, more generally, predictor
values), despite D2D allows such duplicates in the data. This commit
fixes this issue.

There has been a code snippet in a loop over observables which has been
independent of the obervables. This has resulted in unnecessary repeated
assignments inside this snippet. So this commit moves this snippet to a
more suitable location in the body of arInitHierarchical.

Function arInitHierarchical calls function isAlways which by default
rasises a warning when testing a condition like `sym('a')==0`. Such
conditions appear in run of normal operation of arInitHierarchical so
the warning in question is unnecessary. This commit switches it off.

Some of the assertion tests in arCalcHierarchical should be required
only for those observables which have parameters suitable for
hierarchical optimization. Until now, arCalcHierarchical has been
testing such assertions for all observables, thus being overly
restrictive. This commit fixes this issue.

Until now, arCalcHierarchical has been working correctly only for data
series with no missing observations, despite D2D allows such data. This
commit fixes this issue.