…between tests (#1056)

`AccountsManager.deferInitialLoadCatalogsForTesting` is a process-wide mutable
static flag gating whether `AccountsManager.init` spawns its background
`loadCatalogs` Task. The test-safe value is `true` (no background work);
`PalaceTestSetup.bootstrap()` pins it `true` at bundle load. Tests that need the
background load opt IN by setting it `false` in their own setUp.

Two sites left it `false` after running, leaking the unsafe value to the next
test class — which then constructs a real `AccountsManager`, spawning the
1100+-library registry crawl that outlives the test and pollutes whatever runs
next:
  - `AppContainer._resetForTesting()` — runs after EVERY test via the
    singleton-reset observer; its final line set the flag false.
  - `AppContainerResetTests.tearDown()` — explicitly set it false "for
    production semantics."

Both now leave the flag at the test-safe `true`. This is the root cause behind
recurring CI flakes that blocked PRs #1052/#1053/#1054 (none of whose own diffs
were at fault):
  - AccountsManagerCancellationTests.testCancelBackgroundWork_onOptOutInstance
    _isSafeNoOp — opt-out handle non-nil (failed all 3 iterations)
  - CatalogCacheKeyAndIsolationTests — layout-engine-off-main crash
  - BookReturnCleverReauthTests — crash → ** TEST FAILED ** with 0 reported
    failures
  - TokenRefreshOnForeground / TokenRefreshAndRetryQueue — token/network bleed

Regression: AppContainerResetTests.testResetForTesting_disablesBackgroundLoad
Catalogs now asserts the flag is `true` post-reset (was asserting false) — the
mutation that reintroduces the leak fails it.

No production runtime change: the flag is only ever non-default inside an XCTest
process (initializer keys off XCTestConfigurationFilePath); `_resetForTesting()`
already early-returns outside XCTest.

Verified: AppContainerResetTests + AccountsManagerCancellationTests +
CatalogCacheKeyAndIsolationTests + TokenRefreshOnForegroundTests +
TokenRefreshAndRetryQueueTests + BookReturnCleverReauthTests run together →
41 tests, 0 failures, ** TEST SUCCEEDED **.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>