Reduce memory use of GetLevenshteinDistance() by almost 50%

Probably not measurable, but it's also a bit shorter, so probably worth it.
This is a port of my change
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20150713/286948.html
where I explained the idea like so:

  // Although the algorithm is typically described using an m x n
  // array, only one row plus one element are used at a time, so this
  // implementation just keeps one vector for the row.  To update one entry,
  // only the entries to the left, top, and top-left are needed.  The left
  // entry is in row[x-1], the top entry is what's in row[x] from the last
  // iteration, and the top-left entry is stored in previous.

No intended behavior change.

BUG=none

Review URL: https://codereview.chromium.org/1690593002

Cr-Commit-Position: refs/heads/master@{#376015}