harmmean: accept negative and complex values

[dheerchheda77]

Fixes the MATLAB incompatibility reported for harmmean with negative and
complex input.

In MATLAB:

>> harmmean ([1 -1 3])
ans = 9
>> harmmean ([1 1i])
ans = 1.0000 + 1.0000i

In Octave both calls raised "harmmean: X must contain real nonnegative
values."

Cause

The input check at the top of harmmean.m rejected anything that was not
real and nonnegative. The computation below it, n ./ sum (1 ./ x), is
valid for negative and complex values on every code path (default DIM,
"all", DIM, VECDIM, "omitnan"), so only the check had to change.

Changes

• harmmean.m: require only that X is numeric ("harmmean: X must be
  numeric."). Update the docstring, including the notes saying that a 0
  gives 0, which only holds without negative or complex values.
• Tests: replace the error test for [1 -1 3]; add positive tests for
  negative and complex input with default DIM, "all", DIM, VECDIM and
  "omitnan", and for [1 -1] returning Inf.
• doc-cache and NEWS updated.

Not changed

geomean has the same check, but its zero handling assumes real
nonnegative data and I have not confirmed MATLAB's behaviour for negative
input, so it is left out of this PR.

Testing

The expected values in the new tests were checked by hand and with an
independent calculation. Not yet run in Octave locally; relying on CI
for test harmmean. Still to compare against MATLAB: harmmean ([0 1i]).

[dheerchheda77]

Tested locally with Octave 11.3.0 (Windows): test harmmean passes all
28 tests, and harmmean ([1 -1 3]) and harmmean ([1 1i]) now match the
MATLAB results (9 and 1 + 1i). The CI run on this PR started no jobs, so
these local results are the only test run so far.
The doc-cache entry for harmmean was updated by hand; package_texi2cache could not be checked reliably on Windows.