Make mark_reviewed survive reload/rebase only when the file did not actually change.

[melonamin]

UPDATE: merged as #256

Today reviewed state is preserved by path across R reload. That is useful, but it can also keep a file marked reviewed after new commits changed the same path.

Would it make sense to store a per-file fingerprint when the user marks a file reviewed, and on reload keep the reviewed mark only if the fingerprint still matches? Similar to GitHub PR “Viewed” behavior.

Fingerprint probably should be based on the rendered diff state, not only the working-tree file hash, so rebases/new commits that do not change the effective file diff keep the reviewed mark, while real changes clear it.