gather_dict on local error is big bottleneck for large datasets

[daurer]

By default errors for each view are saved at the end of each block of iterations into a dictionary. Those dictionaries are then gathered across all MPI ranks into a global dictionary and might be saved into the .ptyr file if record_local_error is true in the engine params.

For the high-perfomance engines, the dictionary MPI gathering of the errors can be a major bottleneck. In this PR I have made the collection of per-view error metrics optional (using the existing record_local_error parameter) and if not needed, the errors are first reduced on each rank with a subsequent MPI allreduce across all the ranks. This completely removes the bottleneck but still allows collecting the per-view errors if required. By default record_local_error is false.