A long-standing complaint in Review Board 3.0 are the edge cases that

can sometimes be hit with the interdiff filtering algorithm. This has

been difficult to solve, and a great deal of time and investigation has

gone into this for Review Board 4.0. We believe we've improved upon the

algorithm enough to cover all these edge cases, but we need testing.
This change backports the new code to 3.0, making it optional. The new

logic is disabled by default, as we don't want to regress any installs,

but administrators can enable it by setting the following in the

settings_local.py file:

ENABLED_FEATURES = {
    'diffviewer.filter_interdiffs_v2': True,
}


and then either viewing a new interdiff or clearing memcached and

viewing an existing one.
This mechanism allows the new algorithm to be easily turned off if it's

causing trouble.
We're hoping to get people who reported issues in the past to test this

in 3.0.18+, so we can be sure that 4.0 users will have a fully working

interdiff algorithm.
Much of the core algorithmic work was reviewed previously in

https://reviews.reviewboard.org/r/8834/ and commited in cc6a91fa. This

code focuses on bringing in the feature flag support, funneling in the

request object to filter_interdiff_opcodes(), and keying off that

logic by flag.
This is not the final version of this algorithm. Work has been underway

on further improving it to improve upon the output and to add a better

test harness for real-world anonymized diffs, in order to help us build

up a better regression test suite.

Unit tests pass.
Manually tested my interdiff regressions collection with and without

the new algorithm enabled.