Bulk write vector values in distribute_local_to_global.

Bulk write vector values in distribute_local_to_global.

This patch lowers the total number of calls to VecSetValues in
ConstraintMatrix::distribute_local_to_global by writing new vector values to a
temporary array before sending them off to PETSc. There are no notable
performance improvements for the other vector classes (though this certainly
should not hurt).

I used step-40 to check that the performance improved. Here is the (filtered)
output with this patch:

[drwells@archway step-40]$ make run | grep assembly
| assembly                        |         1 |   0.00966s |       8.7% |
| assembly                        |         1 |    0.0122s |       6.4% |
| assembly                        |         1 |    0.0217s |       5.9% |
| assembly                        |         1 |    0.0431s |       5.6% |
| assembly                        |         1 |    0.0848s |       5.6% |
| assembly                        |         1 |     0.173s |       5.6% |
| assembly                        |         1 |     0.351s |       5.5% |
| assembly                        |         1 |     0.729s |       5.5% |

and on master:

[drwells@archway step-40]$ make run | grep assembly
| assembly                        |         1 |   0.00672s |       8.9% |
| assembly                        |         1 |    0.0133s |       7.2% |
| assembly                        |         1 |    0.0258s |       6.9% |
| assembly                        |         1 |    0.0513s |       6.8% |
| assembly                        |         1 |     0.101s |       6.6% |
| assembly                        |         1 |     0.203s |       6.5% |
| assembly                        |         1 |     0.414s |       6.4% |
| assembly                        |         1 |      0.93s |       6.9% |

so it appears that the whole assembly process takes about 10% less time.