Halve the calls of FindCommonElements

[pvc1989]

When I run mpmetis on a 4,000,000-element mesh, the function named FindCommonElements() costs about 11% of the total running time:

  %   cumulative   self              self     total           
 time   seconds   seconds    calls   s/call   s/call  name    
 19.52      0.98     0.98     1510     0.00     0.00  libmetis__Match_SHEM
 18.13      1.89     0.91     1511     0.00     0.00  libmetis__CreateCoarseGraph
 13.15      2.55     0.66        1     0.66     0.66  libmetis__Match_RM
 11.36      3.12     0.57  8000000     0.00     0.00  libmetis__FindCommonElements
  5.58      3.40     0.28        1     0.28     0.85  libmetis__CreateGraphDual

It is called in two loops in CreateGraphDual():

• The only purpose for the first loop is to get the total number of neighbours, i.e. xadj[ne] after L208, which is then used to allocate adjncy in L213.
• The second loop then do the real job, which has already done by the first one.

These two loops could be merged into one, if something supports push_back() like std::vector<int> was introduced.
If so, the total running time might be reduced by about 5%.

[karypis]

Good call.