This commit fixes the bug in phs_symmetrization when it fails to produce ph eigenstates
with PH^2 = +1 when one of the wave functions is mapped exactly onto another one by PH,
as in this case the PH symmetrized wf's from the two are identical. This is the case
with wfs = np.eye(2) and  particle_hole = sigma[1], the second vector returned is an
eigenstate with eigenvalue -1 not +1. In general this can be even worse, if there are
multiple such vector pairs some vectors may not be ph eigenstates at all.

The new code uses an improved version of the matrix square root method from
Applied Mathematics and Computation 234 (2014) 380-384.

Tests are modified to clear up the random unitary generation process (using kwant.rmt).
Tests are now guaranteed to feed orthonormal set of vectors to phs_symmetrization().
Add extra tests for the fully off-diagonal case where the old method fails.

Previously the fact that the positive direction of mode
velocity is determined by the translational symmetry vector
of the system was not made clear in the relevant docstrings.

Closes #143.

There is no point in repeating the docstring in 2 places, though we
leave in the explanation of what this means in the "simplest case", i.e.
lowest modes first, for clarity.

The alleged mode ordering was the opposite of the true ordering.
The true ordering corresponds to "lowest modes first" when there are
no band crossing.

Remove fixes for unsupported versions of numpy, scipy and matplotlib.

Numpy's leastsquares implementation sometimes produces memory errors,
but Scipy's appears to be unaffected (at least for the cases seen
thus far).

Joseph Weston authored Jul 24, 2015 and Christoph Groth committed Jul 28, 2015

Before this commit, if there was a lead for which
the inter-cell hopping matrix was the zero matrix, then
kwant.physics.modes() returned arrays with uninitialized memory.