…rs into ExplicitRouting

PR #2596 added MessageRoutingConvention.PreregisterSenders to make endpoint
policies like UseDurableOutboxOnAllSendingEndpoints see Subscriptions.Any()
on conventionally-routed sender endpoints at Compile() time. The fix worked
for #2588 but introduced a routing-precedence regression: ExplicitRouting
walks Transports.AllEndpoints().Where(ShouldSendMessage), and
ShouldSendMessage just calls Subscriptions.Any(s => s.Matches(messageType)).
Pre-registering subscriptions on the conventional sender now makes
ExplicitRouting fire for handled messages — and since ExplicitRouting is
non-additive, it short-circuits past LocalRouting, so handled messages stop
routing to their local handlers. Explicit publish rules also got duplicated
against the conventional broker exchange.

Eight MessageRoutingTests caught it:
- explicit_rules_always_win.locally_handled_messages_get_overridden_by_routing
- using_additive_local_routing_and_external_conventions.{local_routes_and_broker_conventional_routing_for_handled_messages,explicit_rules_win}
- using_messaging_conventions_for_both_external_and_local.{local_routes_for_handled_messages,multiple_handlers_are_still_routed_to_one_place_in_default_mode,respect_sticky_attributes_but_default_is_still_there_too,respect_sticky_attributes_no_default}
- using_local_routing_disabled_and_external_routing_conventions.explicit_routing_wins_no_matter_what

Fix: tag convention-pre-registered subscriptions with a new internal
Subscription.IsFromConvention flag (via Subscription.ForConventionalType),
and have Endpoint.ShouldSendMessage skip those when answering "is this an
explicit publish rule?". The AllSenders policy still gates on
Subscriptions.Any() (count, not provenance), so #2588's outbox fix keeps
working — verified by re-running Bug_2588 and Bug_2304 reproducers across
RabbitMQ / ASB / SQS / GCP Pub/Sub. The Bug_2588 tests previously
asserted endpoint.Mode via routes.Single() — that worked only because of
this leak. Reworked them to look up the conventional broker sender
directly via IMessageRoutingConvention.DiscoverSenders, which tests the
actual #2588 invariant ("AllSenders policy applied to conventional
sender") without depending on routing precedence.

CI: add CIMessageRouting target (boots RabbitMQ via docker compose, runs
MessageRoutingTests one class at a time) and wire it into the .NET
workflow's CI dependency chain plus a docker compose down step. Without
this, the regression wouldn't have been caught on PRs — MessageRoutingTests
isn't part of any existing per-transport workflow even though the bug
lives in core routing precedence.

Closes the regression introduced by #2596. Refs #2588.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Minor bump for the 5.36 release. Cuts:
- Routing-precedence regression fix introduced by #2596 (PreregisterSenders
  was leaking conventional senders into ExplicitRouting and breaking local
  handler routing for handled message types).
- MetricsTests now wired to the IWolverineObserver pipeline so the
  end-to-end CritterWatch / Hybrid mode tests actually exercise the
  metrics-export path again after d999a2e.
- SQL Server docker image pinned to 2022-latest (2025-latest is broken
  on Linux/ARM hosts).
- New CIMessageRouting target in the .NET workflow so future routing-
  precedence regressions fail fast on PRs.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>