revamp custom_func args

[Mayukhdeb]

all the layer outputs can be stored in a dict instead of a list.

it could look like:

layer_outputs = {
    '__input_image__': input_image_tensor,
    'conv3.branch1': some_tensor,
    'fc': another_tensor
}

and custom_func might look like:

def custom_func(outs):
    loss = outs['conv3.branch1'][:100] + outs['fc'] + torch.var(outs['__input_image__'])
    return -loss