fix(picker): keep branch names out of shared diff-preview cache entries

[max-sixty]

The picker's branch-diff (tab 3) and upstream-diff (tab 4) disk caches are keyed by SHAs only, so branches parked at the same commit — common after wt landed resets merged branches to main's tip — share one entry. The cached value was the finished pane with the row's branch name baked into the empty-diff headline, so every same-SHA row showed the first writer's name: main has no file changes vs main under another branch's row (surfaced as the "preview shows another row's content" oddity while A/B-testing #3439; skim was exonerated — it requests the correct row).

The fix follows the Log cache's existing split: cache only the branch-agnostic data (BranchDiffCacheEntry { body }, UpstreamDiffCacheEntry { ahead, behind, body } — the name-free rendered diff, None when empty) and render the branch-named headline on read, on the hit and miss paths alike. compute_diff_preview now returns Option<String> instead of taking a no_changes_msg, so a branch name can't leak into a cached value again. No key/version bump: old bare-string entries fail struct deserialization, read as a miss (cache::read's documented corrupt-entry semantics), and are recomputed and overwritten in place.

Rendered output is byte-identical in every state, so no snapshot changes.

Testing: regression tests pin the collision for both modes (second same-SHA branch asserted to hit the first's entry and render its own name), a migration test pins old-format-reads-as-miss, and the previously untested empty-diff headlines (clean working tree; upstream ahead/diverged/behind via --allow-empty commits) now have coverage since they're lazily evaluated. Also verified end-to-end by driving the picker in the wt-perf test repo: a poisoned old-format entry self-healed and same-SHA rows render their own names.

This was written by Claude Code on behalf of max