Sum2LoopTrans needs to support `dim` argument.

[arporter]

From my recent work with the NEMO sea-ice, it seems that SUM intrinsics with the optional dim argument still seem to cause the NVIDIA compiler (24.1) problems - I get seg. faults and out-of-memory errors if such reductions are included within KERNELS regions. We have Sym2LoopTrans already but it does not yet support the optional dim argument.