Not updates stacked datasets shape if number of frames is one

The stacking in TrainMatcher does not change the shape of merge_buffer if the number of frames is one.

if merge_data has ones in the leading axes (before staking axis), these ones is broadcasted to any size of of corresponding axis in merge_buffer. And this assignment does not raise exception in this case.

merge_buffer[:, merge_index] = merge_data

mentioned in merge request !72 (merged)

mentioned in commit c3a136fc

closed with merge request !72 (merged)