Fix Segformer test for num_classes 1 with softmax

[sampathweb]

Keras has recently started to emit warnings when softmax has only one class. This PR fixes it by having more than one class at the output. Or we can re-compile the model with BinaryCrossEntropy if we want num_classes to 1.

[innat]

@sampathweb
You're compiling segformer model with BinaryCrossentropy() but the last layer activations is hard coded to softmax. An activation params should be introduced here.