Enable batch_size > 1 for multi_training

[m-klasen]

Hello,
this change allows for training of the temporal transformer with batch sizes of larger than one.
When trying to use larger batch sizes before, it threw a shape mismatch exception in the temporal transformer code section.
I mainly changed the shapes from e.g. [15, 1120, 256] for the configuration with 14 reference frames to
e.g. [2,15, 1120, 256] before the transformer. The transformer input is then e.g. [2, 4200, 10] etc.

[prsbsvrn]

Hi thank you for your code.
Did you also change TransVOD++ code for batch size greater than 1?