Fix a race condition in dbuf_dirty().

We need to wait for pending I/O to finish before trying to clear the
dirty record's override state.

This problem occurs when multiple threads call dbuf_dirty() for the
same dbuf object while it is currently being synced via dmu_sync()
but hasn't reached dmu_sync_done() yet:

  zfs: allocating allocated segment(offset=323629568 size=78336)
  SPLError: 2283:0:(spl-err.c:48:vpanic()) SPL PANIC

Issue openzfs#541

module/zfs/dbuf.c

@@ -1115,6 +1115,10 @@ dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
 		DB_DNODE_EXIT(db);
 
 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
+			/* Wait for pending I/O. */
+			while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC)
+				cv_wait(&db->db_changed, &db->db_mtx);
+
 			/*
 			 * If this buffer has already been written out,
 			 * we now need to reset its state.