[authority_store] Schedule certificate after acquiring tx lock (#4365)

We currently schedule certificate for execution and then acquire tx lock. This opens a relatively small possibility for race condition where execution driver will pick up and lock transaction before we can do that in `persist_certificate_and_lock_shared_objects`. This would not be end of the world, but it would waste time and produce noisy error in logs since execution driver will fail because `assigned_object_versions` was not yet updated.

Simple one line fix is to move scheduling after lock is acquired so that `assigned_object_versions` will be ready for execution driver when it will be able to acquire a lock.

We can't just move `add_pending_certificates` to the end of this function since we want to make sure transaction is scheduled for execution by the time `consensus_message_processed` is updated.

crates/sui-core/src/authority/authority_store.rs

@@ -1110,9 +1110,6 @@ impl<S: Eq + Debug + Serialize + for<'de> Deserialize<'de>> SuiDataStore<S> {
         // Make an iterator to update the last consensus index.
         let index_to_write = std::iter::once((LAST_CONSENSUS_INDEX_ADDR, consensus_index));
 
-        // Schedule the certificate for execution
-        self.add_pending_certificates(vec![(transaction_digest, Some(certificate.clone()))])?;
-
         // Holding _tx_lock avoids the following race:
         // - we check effects_exist, returns false
         // - another task (starting from handle_node_sync_certificate) writes effects,
@@ -1121,6 +1118,9 @@ impl<S: Eq + Debug + Serialize + for<'de> Deserialize<'de>> SuiDataStore<S> {
         // - now it's possible to run a new tx against old versions of the shared objects.
         let _tx_lock = self.acquire_tx_lock(&transaction_digest).await;
 
+        // Schedule the certificate for execution
+        self.add_pending_certificates(vec![(transaction_digest, Some(certificate.clone()))])?;
+
         // Note: if we crash here we are not in an inconsistent state since
         //       it is ok to just update the pending list without updating the sequence.