Learning linear operators: Infinite-dimensional regression as a well-behaved non-compact inverse problem