feat(review): Before/After preview for changed images in code review (#1598) (#1601)

* feat(review): serve Before/After bytes of changed images via /api/review-image (#1598)

Split "which object is each side" from "read it" (resolveDiffSideSources and
its jj/GitButler siblings), so hunk expansion (text) and the image preview
(bytes) share one ref table. Byte readers on both runtimes (git cat-file,
jj file show, capped worktree reads with symlink/realpath containment),
GitHub/GitLab base64 byte fetchers, and one shared request handler in
packages/shared/review-image.ts (eligibility, caps, magic-byte sniffing,
header dimensions, error mapping, nosniff + sandbox CSP + CORP headers).
The Bun and Pi servers only supply where the current mode reads a side,
and advertise imagePreviewSupported on every diff payload.

* feat(review): Before/After preview for changed images in code review (#1598)

ImageDiffPreview replaces the binary notice for a hunkless image chunk in
the single-file and all-files views when the server advertises
imagePreviewSupported: Before | After panes (one pane for added/deleted,
"Contents unchanged" for identical bytes), dimensions and size deltas, a
theme-token checkerboard, stacking for narrow/compact layouts and tall
images. Lazy (IntersectionObserver), abortable, object-URL LRU per
snapshot; SVG only via <img src=blob:>. The all-files view takes it through
a render prop only the review app sets, so guide hosts keep the notice.
Regenerated the guides.show viewer manifest; docs in AGENTS.md.

* fix(review): address #1601 review: abortable image reads, bounded client cache, dotted file names

- Server read limiter takes the request abort signal (Bun req.signal, Pi
  response close): queued reads for cards scrolled past are dropped before
  they run, and no gh/glab call starts for a request that is gone.
- Client cache caps total bytes (128 MB) besides entries, never revokes an
  object URL a mounted <img> still uses, and ignores a response that lands
  after the page moved to another snapshot.
- validateFilePath rejects ".." path segments (and absolute paths), not any
  ".." substring, so a..b.png previews; traversal stays refused, tested on
  /api/review-image and /api/file-content in both runtimes.
- Only an HTTP 404 means a PR file is missing; a missing CLI is a 502.
- All-files panes size to the image (same max); misplaced comments moved;
  AGENTS.md documents freshness, aborts and the cache bounds.

* fix(review): refuse dot-only path segments; keep the first URL for overlapping image fetches

- validateFilePath rejects any segment made only of dots, not just "..":
  "..." is Perforce's recursive wildcard and /api/file-content passes the
  path to p4 print in P4 sessions (a whole-subtree read). a..b.png stays valid.
- fetchReviewImage: when two fetches for the same side overlap, the later
  response revokes its own URL and returns the cached entry instead of
  revoking a URL a card may already be showing.