Fix PublishingRelay misrouting default-tenant events under conjoined tenancy

[svenclaesson]

Closes #2675

Summary

Under conjoined tenancy, PublishingRelay's default-tenant fallthrough let WolverineSubscriptionRunner's bus.TenantId (set to operations.Database.Identifier) leak onto the outbound envelope. Default-tenant events were dispatched under the database identifier (e.g. "Main") instead of *DEFAULT*, the handler chain ran under the wrong tenant scope (or silently no-op'd), but Marten still advanced the subscription sequence because publishing succeeded.

This change gates the override on Events.TenancyStyle == Conjoined, so only the broken case is touched. Non-conjoined stores keep the existing fallthrough exactly, preserving setups that legitimately use Database.Identifier as the message tenant (e.g. per-tenant ancillary stores with TenancyStyle.Single).

Compatibility

Setup	Pre-fix Envelope.TenantId	Post-fix Envelope.TenantId
No tenancy (default Marten store)	Database.Identifier	Database.Identifier (unchanged)
Conjoined, default-tenant event	Database.Identifier (the bug)	*DEFAULT* (correct)
Conjoined, named-tenant event	"one"	"one" (unchanged)
MultiTenantedDatabases(...AddSingleTenantDatabase("tenant1"...))	"tenant1"	"tenant1" (unchanged)
Per-tenant ancillary stores with custom Database.Identifier (TenancyStyle.Single)	the custom identifier	the custom identifier (unchanged)

The only behavioural delta is the targeted bug case. Every other row keeps its pre-fix behaviour because the override only activates in conjoined mode.

Tests

Two regression tests added in MartenSubscriptionTests:

1. carry_default_tenant_id_through_under_conjoined_tenancy — sets up a conjoined store, publishes a default-tenant event, asserts Envelope.TenantId == StorageConstants.DefaultTenantId. Reproduces the bug on main (asserts "*DEFAULT*", gets "Main"); passes after the fix.
2. non_conjoined_store_preserves_legacy_default_tenant_fallthrough — companion test using a plain non-conjoined Marten store; asserts the relay still falls through to the bus context's tenant id, guarding the per-tenant-ancillary-store pattern against accidental over-application. Passes both before and after.

MartenSubscriptionTests runs green on net8/net9/net10 (12/12).

Test plan

• Reproduce the bug with a failing test on main
• Apply fix; regression test passes
• Companion test confirms non-conjoined fallthrough preserved
• Full MartenSubscriptionTests suite green on net8.0, net9.0, net10.0
• Full MartenTests suite (rerun in progress; previously: 433/439 with 6 unrelated infrastructure flakes — postgres connection pool exhaustion + agent-distribution timeouts; none touch PublishingRelay)

See #2675 for the full root-cause trace through WolverineSubscriptionRunner → PublishingRelay → MessageBus, and a side-by-side with the Polecat sibling that already handles tenancy correctly.

[svenclaesson]

Broader MartenTests rerun completed: 434/439 pass on net9.0 with this fix applied.

The 5 failures are pre-existing infrastructure flakes unrelated to PublishingRelay:

Test	Failure mode
Persistence.marten_durability_end_to_end.sending_recovered_messages_when_sender_starts_up	message recovery timing
Distribution.basic_agent_mechanics_versioned_composition.start_with_multiple_databases_on_one_single_node	agent-distribution timeout
Distribution.basic_agent_mechanics_versioned_composition.spread_databases_out_via_host	agent-distribution timeout
Saga.When_starting_an_order.should_not_be_completed	Failed to setup resource Main of type WeaselDatabase
Saga.not_found_usage.try_to_call_handle_on_already_expired_invitation	Npgsql.PostgresException 53300: sorry, too many clients already

Verified by inspection that none reference PublishingRelay, PublishEventsToWolverine, or SubscribeToEvents. A baseline run on main (without this fix) showed 433/439 with the same failure modes — i.e. the fix didn't introduce any regression and slightly improved the count (one fewer flake hit).

[jeremydmiller]

@svenclaesson I'm going to change this one just a wee bit to try to remove the runtime logic and the downcasting tomorrow, but the test will go in as is.

[jeremydmiller]

@svenclaesson I did this just a little bit differently, but you're listed as the co-author of the new PR