New implementation of sprandsym (please test)

[Jordi Gutiérrez Hermoso]

Following Rik's request to fix the sprand functions, I attempted a new
implementation of sprandsym. I hope it's not too convoluted or
difficult to understand:

    http://hg.savannah.gnu.org/hgweb/octave/rev/abf1e00111dd

All tests now pass, including the explicit density one. Profilng it,
it seems to perform only slightly slower than the previous version,
due to slowness on randperm for large inputs. I am considering some
simple optimisations for randperm to fix this.

HTH,
- Jordi G. H.