Fix 2x discrepancy between CPU and GPU for TKE periodic BC node rhs

coeffs.

Now, comparing log files for CPU vs GPU runs (for just 1 MPI rank
in each case, to remove that as a variable), the TKE residuals
match very well, and the 'Mean system norm' matches in 9 decimal
places after 10 time steps.

For follow-on work, we do need to convert to using the ngp version
of periodic_field_update which should help performance.

src/ngp_algorithms/TKEWallFuncAlgDriver.C

@@ -68,6 +68,9 @@ void TKEWallFuncAlgDriver::post_work()
     realm_.periodic_field_update(bcNodalTkeField, nComp, bypassFieldCheck);
   }
 
+  ngpBcNodalTke.modify_on_host();
+  ngpBcNodalTke.sync_to_device();
+
   // Normalize the computed BC TKE at integration points with assembled wall
   // area and assign it to TKE and TKE BC fields on this sideset for use in the
   // next solve.