Allow modules to pass lists of outputs in case of parallelization across several GPUs or gradient accumulation

[damro]

Optimizer needs to be able to collect losses across multiple batches or GPUs. Therefore the API between modules should allow some form of lists of outputs and losses