Fix MSVC+CUDA issues.

Darn MSVC+CUDA gets confused for diagonal and transpose again, not able to match out-of-line definitions with the corresponding declarations. Removed Modified internal typedefs and just use the true type to get around this.

MSVC also complained about not passing enough arguments to function-like macro, and about invalid friend declarations. Removed unused macro argument, and explicitly specified friend classes to get around these.