Fix test-fixture reset routing, outbox restart overlap, and test-harness consumption/disposal bugs

[Vulthil]

Summary

• BaseIntegrationTestCase's post-test reset always ran against the shared class-fixture factory, even
  when a test overrode CreateFactory() to run on a WithWebHostBuilder(...)-derived host. That silently
  built a second, unused host and paused the wrong one's background services while Respawn wiped the
  shared database. Reset now targets whichever factory the test actually ran on, and is skipped entirely
  when a test never built a host.
• Giving the reset a real, cancellable timeout (instead of an internal-only bound) opened a window where
  OutboxBackgroundService.StartAsync could start a second relay loop before a slow-to-stop previous one
  had actually finished. StartAsync now waits for an incomplete previous execute task before starting a
  new one.
• The in-memory test harness recorded a request as "consumed" even when the request consumer threw,
  because RecordConsumed ran before the consumer's await instead of after (already correct for one-way
  consumers).
• BaseUnitTestCase now disposes a created Target on teardown if it's disposable.
• Docs/XML-doc corrections: AddHttpMock's "configure up front" language contradicted the documented
  per-test reset behavior; ITestHarness's singleton-lifetime capture leak across tests sharing one host
  is now documented with a Clear()-from-Initialize() pattern.
• Re-verified several previously reported fixture issues (container-collection mutation during parallel
  iteration, a teardown-skipping initialization guard, lazy scope/client thread-safety, a Cosmos
  readiness wait "falling through") against current code and found them already fixed - no change needed.

Verification

• Full solution build: clean.
• Unit suites (Vulthil.SharedKernel.Outbox.Tests, Vulthil.Messaging.TestHarness.Tests,
  Vulthil.SharedKernel.Tests): all passing, net9.0 and net10.0.
• Full Vulthil.IntegrationTests suite against real Postgres/MySQL/Cosmos containers: all passing,
  net9.0 and net10.0 - confirms no regression to existing integration coverage.
• Each new/extended test was confirmed to fail against the pre-fix code and pass against the fix.

Backport to v1.0: no - test-infrastructure behavior changes; ships with the next minor.

---

Amendment: UseReal/UseRealFor now register the real instance as a lazy factory instead of constructing it
immediately - it is built the first time it's actually resolved (directly or as a dependency), so a dependency
registered afterward but before that resolution is now visible to it (previously it would have been auto-mocked).
BaseUnitTestCase.Dispose() now also disposes every disposable instance the auto-mocker holds (Use() instances,
resolved UseReal/UseRealFor instances, auto-generated dependency mocks) via AutoMocker.AsDisposable(), in
addition to the existing Target disposal - Target is still disposed first. Only synchronous IDisposable is
covered; an IAsyncDisposable-only registration still needs a manual override. These are the two behavior changes
reviewers should look for: construction timing for UseReal/UseRealFor, and new automatic disposal of anything
the auto-mocker holds.