20161109 bugfix in models.transition.remove_rows

[moritz-kirmse-forcity]

Correction for the following problem:

error handling is too strict:
whith accounting_column != None, the function won't allow removing an amount len(data) < nrows < data[accounting_columns].sum(), even if this is possible.

[moritz-kirmse-forcity]

Checking why tests have failed.
Failed tests might be related to the random sampling in transition module. If the algorithm does not converge after 50 iterations an incorrect result could be returned depending on the random seed.

[moritz-kirmse-forcity]

I found 2 issues with this PR:

1. deprecated method Index.diff has been removed in pandas 0.19, breaking some of the urbansim code
2. I had some failed tests where the sampling in remove_rows does not return the correct total with respect to accounting_column. -> opening an issue for that No feedback if iteration does not reach exact result in utils.sampling.sample_rows #178:

However, the bugfix remains valid even if the tests fail.

[moritz-kirmse-forcity]

updated the tests for transition, the tests pass on local system with pandas 0.18.1
There is still a workaround in line 154 due to misbehavior of the random permutation. This shall be updated when issue #178 is solved

[coveralls]

Coverage increased (+0.006%) to 94.179% when pulling fa1fcb7 on moritz-kirmse-forcity:20161109_fix_remove into caa6ab8 on UDST:master.