Occlusion: No multiply attr_output_fn with output

[chr5tphr]

• do not multiply the output of attr_output_fn with the model's output,
  as this is neither intuitive nor flexible
• instead, only sum over the attr_output_fn output to compute the
  batch-wise scores
• in share/example/feed_forward.py: specify an attr_output function instead
  of a one-hot tensor, which uses only the output of the true label
  index, thus retaining the previous behavior and demonstration how to
  use the output function
• in docs/source/how-to/use-attributors.rst: use a lambda expression for
  the occlusion attributor to demonstrate how to use score only for a
  certain index with the new interface