GooFit cleanup, consisting of:

• Using standard terminology consistently for CUDA (closes Clean up terminology #57)
• Removing the fakecuda directory and hacks associated with it
• Removing devcomplex, and moving to using thrust::complex everywhere (closes Complex numbers #58)
• Removing using namespace std from all header files
• Use #pragma once in all headers
• Removing includes from header files, using forward declarations instead, and includes in src: Faster compilation, especially for PDFs
• Removing many warnings about initialization order, using C++11 style in more places
• Use Arch 3.5+ features through the compiler, allowing mixed targets, removed flag
• Expand __ldg to support non-intrinsics, using generics mini-library
• Dropping many of the odd help print defines that were untested and often invalid
• Removing the ALL_CAPS math defines; CUDA already selects the correct one for float/double
• Several classes moved to header (Variable, FitControl)
• Use submodules instead of DownloadProject, automated in CMake, moved libraries to /extern
• Examples now give an error if the fit fails (FitManager can be used as int or bool), better logs for RunAll.py
• Removed a couple of headers for root stuff that managed to survive
• Added warnings to --goofit-details if optimizations are missed, added better optimization for host code on CUDA
• Made dependency on rang simpler, GooFit::red == rang::fg::red
• Unrecommended, but in-source builds are supported again

scripts/ModernizeGooFit.py knows about most of the new changes and can assist in converting old code.