Less broadcast for numpy exercise

Currently the indexing in the numpy example and exercise revolve a lot around broadcasting.

This is not necessary very useful. Other kind of indexing, like selection with boolean mask or range selection may be more interesting.