Adding a "distributed" option to the LogReport class.

[linshokaku]

During distributed learning, the LogReport class cannot aggregate reports from all processes. This should be fixed because it can cause bugs when using extensions or options that depend on the values in the logs. For example, there are issues when using EarlyStoppingTrigger or ReduceLROnPlateau.

To address this problem, we need to modify the LogReport class so that it gathers the summary objects from all processes at the trigger point and recalculates the averages. We should also add an "writer_rank" option to ensure that only one process creates the log file.