Bug on edit distance align

[yzhaoinuw]

This pull request attempts to fix the issue raised by @yzhangcs's comment as a follow-up to #3017. In short, when the substitution cost is greater than 2, the path generated by _edit_dist_backtrace() still includes substitution. This is incorrect because a deletion followed by an insertion incur a cost of 2, which is lower than substitution, and thus should be preferred over substitution.

In the following test case due to @yzhangcs, the current _edit_dist_backtrace() gives the following output

from nltk.metrics.distance import edit_distance_align

>>> edit_distance_align('a', 'b', float('inf')) 
[(0, 0), (1, 1)]

After the fix it correctly outputs [(0, 0), (0, 1), (1, 1)]. I tried to add the test case provided by @yzhangcs, but didn't figure out how. Could the maintainers of NLTK look into this and add this test case?