- The selection criteria for the two roots of each second degree equation
  has been revert, as in the original script. Another test for null roots
  has been added few lines after.

  An additional test for null dt (or too small value), has been added,
  and Ltol (weighted by h, the cell characteristic length) is used as
  scaling for dt.

  The computation of the characteristic length h is straightforward, since
  it uses directly the determinant of the matrix A, already known (detA
  is the area of the triangle, therefore h is approximated by its square
  root).

  We can avoid to call tsearch and to compute the 3-by-3 matrices A and B
  at each time step. A test has been added to check when the path is going
  to leave the triangle.

  Last, dLtol should be also weighted by h, in order to get a similar
  distribution of points, whatever the mesh resolution is (since the role
  of dLtol is about the path curvature). But this is done only when the
  current triangle doesn't lead to a stagnation point.

  In some cases, I found that the L increment, that is to say 'a*dt+b', is
  equal to (or close to) the null vector, although dt is not small in
  comparison to unity. This is not exactly a stagnation zone (already
  checked at the end of the routine, by comparing the scalar value of dt_c
  with du_max), but it is due to the presence of a stagnation point, or
  even a stagnation line inside the triangle.
  All these pathological cases have been (hopefully) fixed.

