Fix empty array edge cases in coxphfit - #488

MATLAB's result for no data comes from a singular matrix inverse (covb -Inf on the diagonal, se Inf, p and LikelihoodRatioTestP 1), and its "rank deficient" warning names the wrong cause. We don't replicate that, and MATLAB's own mnrfit, mvregress and robustfit error on empty input. Keep coxphfit erroring when no complete observation remains; a clearer message for empty input is welcome. Also, warning ("off", ...) takes an identifier, not the message text, so that test cannot work.

updated @pr0m1th3as
can you review again, thank you.

Please revert the new check on T: for a 3-column T it replaces the clearer existing message about (start, stop] intervals. Dropping the two ! isempty tests is enough.

The empty check runs after the NaN rows are removed, so coxphfit ([NaN; NaN], [1; 2]) now reports empty input instead of "no complete observations remain after removing NaNs". Check straight after the row count check instead, and leave the NaN branch as it was:

  if (rows (X) != n)
    error ("coxphfit: T must have one element for each row of X.");
  endif
  if (n == 0)
    error ("coxphfit: X and T must contain at least one observation.");
  endif

Keep X and T in upper case, give the %!error tests the full message, and write zeros (0, 3) with spaces.

updated @pr0m1th3as