Remove function pointer based matrix accessor

As it turns out, function pointers in CUDA code are a bad idea. They are used in the MatrixAccessor class of our  SPOOF-CUDA operators to access dense or sparse side inputs at runtime. The idea was to set the access method once in the beginning of a kernel to avoid a condition on every access but tests have shown that an "if dense do A else do B" works just fine. Resorting to this atm and leave a more elaborate yet quick solution as a ToDo.