GH-4556: apply an ISendMyself published from a projection side effect

[erdtsieck]

Closes #4556.

Problem

MartenToWolverineMessageBatch always passes DeliveryOptions (at least the tenant id) to PublishAsync. MessageBus.PublishAsync only lets an ISendMyself apply itself when options == null. So a ToWebSocketGroup(), ToEndpoint(), DelayedFor(), ... published from RaiseSideEffects gets routed as the wrapper type and ends up as NoRoutes.

Fix

The batch now spots an ISendMyself and applies it against its own ProjectionSideEffectContext:

• a MessageContext enlisted in the same Marten session through MartenEnvelopeTransaction, with the batch's storage, so durable sends still commit with the projection
• flushed in AfterCommitAsync after the batch's own context
• it overrides TrackEnvelopeCorrelation, the point every outgoing envelope passes through (publish, EndpointFor, topics), to stamp the side effect's tenant, correlation id, causation id and headers. Anything the message set on its own options wins.

I used a context per message instead of changing TenantId/CorrelationId on the shared batch context for the duration of ApplyAsync, because slices can publish concurrently.

I left the options == null guard in MessageBus alone, since it protects against the TimeoutMessage loop.

Tests

Bug_4556_projection_side_effects_honor_send_myself covers ToEndpoint() from an async projection, with metadata (correlation, causation, custom header asserted on the received envelope) and without. Both failed on main with NoRoutes for RoutedToEndpointMessage<SendMyselfCommand>.

Also still green locally: Bug_2545, Bug_4268, Bug_2382, Bug_2887, end_to_end_publish_messages_through_marten_to_wolverine, multi_tenancy_queue_usage, and the MartenTests.AncillaryStores namespace (46 tests).

[jeremydmiller]

Thank you for this @erdtsieck — the diagnosis was exactly right, and the ProjectionSideEffectContext
approach (stamping the metadata in TrackEnvelopeCorrelation, the one point every outgoing envelope
passes through, rather than mapping it onto a single DeliveryOptions) is the design we kept.

Closing in favour of #4575, which carries your commit authorship and adds the review follow-up:

• The correlation id went null. MessageMetadata.CorrelationId is plain null unless the
  projection set one, so taking it over unconditionally stamped null — stripping the correlation id
  that every other side-effect message out of the same batch carries. Reproduced with a plain message
  published from the same slice as the control; it now falls back to the batch context's id.
• _sendsThemselves was never cleared. Marten only discards its cached message batch after a
  successful commit (SaveChangesAsync doesn't reset it in a finally), so a session retrying a
  failed commit gets the same batch back and would accumulate — and re-flush — every context from
  every failed attempt for as long as it lived. Now cleared on flush, lazily allocated, and the
  Add/iterate locking is symmetric.
• Polecat and Fisher had the identical defect. PolecatToWolverineMessageBatch and
  FisherToWolverineMessageBatch were byte-identical to Marten's down to the doc comments, and both
  still dropped every ISendMyself as NoRoutes — confirmed with a red baseline on all three stores.
  Rather than triplicate the fix, the behaviour moved into a store-agnostic
  ProjectionSideEffectSink that all three bridges delegate to.
• ToTopic() was the one wrapper that still wouldn't have worked. BroadcastToTopicAsync never
  called TrackEnvelopeCorrelation at all, so it carried no causation id or custom headers (and no
  Source/CorrelationId/Store either — for every topic broadcast, not just side effects). Fixed
  in its own commit there so it can be taken separately.

One note on the test: the schema clean ran after StartAsync(), which drops mt_event_progression
out from under a daemon that is already polling it — a reliable source of CI flakiness. #4575 drops
the schema before the host starts instead.