Re-enable an important optimization in the Myers diff algorithm.
There's an optimization we turned off many years ago in MyersDiffer

that results in excessively long diff times for very large files. This

was the source of some huge CPU spikes and memory usage, especially if

users dogpiled on the review request.
We turned this off in the early days of Review Board, and the commit

message was vague about why. It seems that it resulted in a breakage

with some diffs. I have not been able to reproduce this, and this part

of the algorithm matches GNU diff's perfectly. I know we've made other

fixes to the differ since then, so most likely, the breakage was due to

one of those.
To ensure compatibility with older diffs, I've bumped the diff compat

version. If it turns out that this does break things, we can easily

revert it, and test manually with any diffs by setting the stored compat

version in the database per-diffset.
This gets some large, more insane diffs from 1-2 minutes down to around

10 seconds.

Unit tests pass.
Tested with some large diffs that went down the optimized code path. Didn't

see any problems.