Make PSPNet Fully-convolutional

[mrlzla]

There was no way to use PSPNet as fully convolutional network because we had hardcoded AveragePooling builded in interp_block:

if input_shape == (473, 473):
        kernel_strides_map = {1: 60,
                              2: 30,
                              3: 20,
                              6: 10}
    elif input_shape == (713, 713):
        kernel_strides_map = {1: 90,
                              2: 45,
                              3: 30,
                              6: 15}
    else:
        print("Pooling parameters for input shape ",
              input_shape, " are not defined.")
        exit(1)
kernel = (kernel_strides_map[level], kernel_strides_map[level])
strides = (kernel_strides_map[level], kernel_strides_map[level])
prev_layer = AveragePooling2D(kernel, strides=strides)(prev_layer)

So we were able to use the network only with static shape. That is not flexible because:

1. We could only train the models that have input_shape either (473, 473) or (713, 713)
2. We could only train the models with one shape ( f.e. we couldn't feed two images with shape (473, 473) and (713, 713) during one train process).
   This two restrictions were fixed with this pull_request.