fix(frontend/library): unify schedule cache invalidation across all 4 affected query keys (#13204)

## Why

Followup to **#13202** (the autopilot UX + unified Scheduled page PR I
just merged).

During a critical post-merge audit prompted by the user's "be critical,
test all corner cases" prompt, I found that the unified
`/library/followups` page + the new "Autopilot library" briefing pill
BOTH read from the user-wide `/api/v1/schedules` endpoint, while the
existing per-agent schedule UI (`/library/agents/{id}`) reads from the
per-graph `/api/v1/graphs/{id}/schedules` endpoint. These are different
React Query keys backed by different endpoints, but they cache
**overlapping data** — every graph schedule lands in both.

Every mutation site (8 total — 1 from #13202, 7 pre-existing) was only
invalidating ONE of the affected keys. **Native /pr-test surfaced a 4th
stale consumer**: the `/library` fleet-summary "Scheduled" count derives
from `agent.is_scheduled` on the library-agents query, which no mutation
site ever invalidated.

Net effect: every schedule create / edit / delete leaves at least one
consumer (and often two) with stale cache until manual reload.

## What

**Single shared invalidator** wired into all 8 mutation sites,
invalidating **4 query keys**:

| Query key | Consumer |
|---|---|
| `/api/v1/schedules` (user-wide) | unified `/library/followups` +
Autopilot library pill |
| `/api/v1/copilot/followups` | copilot followup list inside the unified
page |
| `/api/v1/graphs/{id}/schedules` (per-graph) | agent detail sidebar +
selected-schedule view |
| `/api/library/agents` (partial-key match) | `/library` fleet-summary
"Scheduled" tab count, all variants (infinite, search, filter) |

| Mutation site | Old invalidation | Now |
|---|---|---|
| `useGraphScheduleListItem` (unified page delete — #13202) | user-wide
only | ALL four |
| `useFollowupListItem` (unified page followup delete — #13202) |
copilot-followups only | ALL four |
| `useScheduleDetailHeader` (agent detail delete) | per-graph only | ALL
four |
| `useSelectedScheduleActions` (selected-schedule delete) | per-graph
only | ALL four |
| `ScheduleActionsDropdown` (sidebar dropdown delete) | per-graph only |
ALL four |
| `useScheduleAgentModal` (create from agent detail) | per-graph only |
ALL four |
| `useEditScheduleModal` (PATCH from agent detail) | per-graph only |
ALL four |
| `useCronSchedulerDialog` (create from agent **builder**) | **NONE** 🐛
| ALL four |

That last one is the worst pre-existing bug: scheduling from the builder
didn't refresh the agent detail page either — completely independent of
my PR.

## How

New helper at
`autogpt_platform/frontend/src/services/schedules/invalidate-schedules.ts`:

```ts
export function invalidateAllScheduleQueries(
  queryClient: QueryClient,
  graphId?: string,
) {
  queryClient.invalidateQueries({ queryKey: getGetV1ListExecutionSchedulesForAUserQueryKey() });
  queryClient.invalidateQueries({ queryKey: getListCopilotFollowupSchedulesQueryKey() });
  queryClient.invalidateQueries({ queryKey: ["/api/library/agents"] }); // partial-key — hits every variant
  if (graphId) {
    queryClient.invalidateQueries({ queryKey: getGetV1ListExecutionSchedulesForAGraphQueryKey(graphId) });
  }
}
```

Every mutation site now calls this single helper instead of
cherry-picking one key. `graphId` is optional because copilot followup
mutations don't bind to a specific graph.

The library-agents key (`["/api/library/agents"]`) is matched as a
**partial prefix** — Tanstack Query's default behaviour — so it catches
the bare query AND every variant with params (search, filter,
pagination, plus the infinite-query form used by `/library`).

## Test coverage

- **Helper unit tests**: `invalidate-schedules.test.ts` — 2 tests
verifying both branches (with + without graphId) and asserting the exact
set of 3 or 4 invalidated query keys.
- **5 new hook test files** with renderHook + MSW, one per modified hook
— each spies on `invalidateAllScheduleQueries` to assert it's called
with the right `graphId` on the success path:
  - `useScheduleDetailHeader.test.tsx` (3 tests)
  - `useSelectedScheduleActions.test.tsx` (3 tests)
  - `useScheduleAgentModal.test.tsx` (2 tests)
- `useEditScheduleModal.test.tsx` (2 tests, with restored
`globalThis.fetch` per CodeRabbit nit)
- `useCronSchedulerDialog.test.tsx` (2 tests — explicit regression for
the previously-zero-invalidation bug)
- **Native /pr-test**: 4 scenarios pass on a local poetry/pnpm stack
with real backend + real browser interactions. See [test report
comment](https://github.com/Significant-Gravitas/AutoGPT/pull/13204#issuecomment-4527332405)
for screenshots.
- **Full vitest suite**: 2802 tests pass locally; codecov patch above
the 70% target.

## Out of scope

- No backend changes — pure client-side cache hygiene.
- The execution `runs` query (`getGetV1ListGraphExecutionsQueryKey`) is
intentionally NOT included in the helper since not every schedule
mutation creates a run; only the "Run now" path needs it and that
already invalidates it inline.

## Checklist

- [x] Code follows project style guidelines
- [x] Self-review performed
- [x] Helper unit-tested + every changed hook unit-tested
- [x] Native /pr-test confirms cross-page propagation across 4 scenarios
- [x] No new warnings
- [x] Pre-existing behavior preserved (mutations still invalidate the
keys they used to + the missing ones)
- [x] CI green