…cache (#3678)

* fix(general): surface original commit-phase-2 errors to transaction caller

Tx#executeCommit2 used to wrap any phase-2 participant failure in a
generic FinalizationError carrying only the participant name. The
original error (e.g. a StatusResponseError returned by a remote device
declining a write) was logged but never reached the caller, so
ClientNode setStateOf / agent.state / endpoint.set callers had no way
to detect the device's status code.

throwIfErrored now rethrows the original error when a single participant
fails, and aggregates with MatterAggregateError when multiple do. The
rollback path's mislabeled "in commit phase 2" message is corrected.

Adds direct unit tests for the new contract (TransactionTest) and
end-to-end ClientNode tests covering single rejection, partial multi-
attribute rejection, and full multi-attribute rejection across two
clusters, plus a server-side regression test for the preCommit path.

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

* feat(node): roll back local client cache on declined remote writes

After PR #3676 the original StatusResponseError reaches the caller of
ClientNode setStateOf / agent.state / endpoint.set, but the local
client-side cache still held the value the user attempted to write.
Subscriptions only deliver server-side changes, so the divergence
persisted until restart.

RemoteWriteParticipant now tracks a per-cluster snapshot
(previousValues + writtenValues + compensator) per set() call. On
commit2 failure the new RemoteWriter onFailure callback delivers
per-attribute statuses; the participant groups failures by
endpoint+cluster and invokes the compensator with the failed ids.

DatasourceCache implements the compensator: it restores the pre-write
value via the existing externalSet path (so persistence + listeners
fire as if a subscription update arrived) but only when:
  - a baseline previous value was captured for the key, AND
  - the current local value still equals what was just written
    (skip restoration if a concurrent subscription update changed it).

Adds three end-to-end tests covering single decline, partial decline
across two clusters, and full decline.  The tests fail when the
compensate body is disabled, so they verify compensation rather than
relying on subscription delivery.

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

* refactor(node): use reference identity for write-decline compensation guard

Datasource keeps user-written values by reference and integrateExternalChange
replaces a reference only when a subscription actually delivers a changed
value (its own no-op deep-equal filter strips identical re-deliveries).  The
guard in DatasourceCache.compensate therefore only needs to check `current[id]
!== writtenValues[id]` instead of running a full deep-equal on every key — the
ref check captures real concurrent updates without the extra walk.

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

---------

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