Unify NormalOperator Handling

[tknopp]

Our handling of the normal operator is currently not satisfying. Some methods have the AᴴA argument, CGNR has no way to pass the normal operator but it generates it internally. There it uses our "normalOperator" infrastructure for lazy generation of the normal op.

What we want:

• Let the user create the normal op. Otherwise create a trivial one
• The trivial one needs to be lazy. we never want to calculate the full matrix product
• Potentially we can get rid of the LinearOperator / SparsityOperator dependency.

[tknopp]

solved in https://github.com/tknopp/RegularizedLeastSquares.jl/commit/8430973d3c0d26474b8af3ff76b393e52b5d426b We got rid of SparsityOperator. LinearOperator is still there but can be probably removed with not so much effort (opEye is the only remaining function being used)