Adopt Vue composables for stateful non-rendering logic, starting with TemplateSection's scroll-arrows

Why now

Found while investigating a recurring test-coverage gap across #126/#127/#130 (tracked separately in #133): the same file keeps absorbing new interactive elements and derived state without ever being split up. Composables are the piece "extract it" doesn't cover on its own — component extraction splits render output, composables split stateful logic with no direct template dependency. This project is small (5 Vue files, ~1000 lines total) and young, so establishing the convention now is cheap; it gets more expensive the more code lands on the current shape.

PR 1 — proof of concept: TemplateSection.vue's scroll-arrows cluster

list (element ref) + canScrollLeft/canScrollRight + updateArrows() + scrollByStep() + the onMounted/onUnmounted wiring is a self-contained "scrollable region with arrow buttons" state machine — reusable, no dependency on other in-flight work, and the same region already has an open a11y gap (#59: no aria-label identifying it as a scrollable region) worth closing while in there.

Check @vueuse/core before hand-rolling this one. Verified directly against its shipped type declarations: useScroll(element) already returns arrivedState.{left,right,top,bottom} (the inverse of canScrollLeft/canScrollRight) plus reactive x/y, and handles the listener attach/detach and resize-observation internally. The extraction should be a thin wrapper over useScroll, not new hand-rolled logic — this is the same "reuse before you write" ladder AGENTS.md already applies to @nextcloud/* packages and @nextcloud/vue components, extended to general-purpose Vue utilities. @vueuse/core@14.4.0's only peer dependency is vue@^3.5.0, which this app already satisfies.

Conventions to establish with this PR, not invent per-composable later:

1. Check @vueuse/core first. Before hand-rolling a composable, check whether it already exists there (as with useScroll above). Prefer it the same way an existing @nextcloud/vue component or design token is preferred over a hand-rolled equivalent.
2. Return individual refs, not a reactive() object. A composable that returns reactive({...}) breaks reactivity silently if the caller destructures it. Return individual refs (or toRefs() on an internal reactive object) so destructuring is always safe.
3. A shared test helper for composables using lifecycle hooks. onMounted/onUnmounted throw outside an active component instance, so testing needs a small reusable helper — a trivial host component whose setup() just returns the composable — living in src/test-utils/, written once here rather than reinvented slightly differently by the second and third composable.
4. Same new-unit-plus-sibling-spec rule as src/utils//src/components/ today. src/composables/useScrollArrows.ts lands with src/composables/useScrollArrows.spec.ts in the same commit — this is also where the src/composables/ directory convention itself gets established.

PR 2 — once PR 1 is proven, extend deliberately to the rest

Not opportunistic, not "whoever next touches that region" — a committed second PR, once PR 1's shape has held up in review and in practice. Targets:

• OfficeOverview.vue: the create-from-template dialog cluster (showCreateDialog, newFileName, pendingCreator, pendingTemplate, creating, createError, createInput)
• OfficeOverview.vue: the view-mode toggle + persistence cluster (viewMode, toggleViewMode(), getOverviewGridView/setOverviewGridView)
• OfficeOverview.vue: the active-creator/routing-sync cluster landing via #130 (activeCreator, routeCreatorId, the URL-sync watch)

Each of these is its own commit within PR 2, one concern per commit, same as always — but landing as one deliberate pass rather than trickling in over unrelated future feature work.

After both PRs: documenting the pattern

Once PR 2 lands and the model has held up across all four clusters — not before — #139 writes it into AGENTS.md as an established pattern. Kept as a separate issue rather than folded into this one, so this issue stays scoped to the code work.