Fixes for TX1 64-bit OS + CUDA CMake Fix

FEAT Function to return allocated bytes

FEAT Scalar return for dot

Also remove boost inline for older CUDA versions

Access to InteropManager has to go through device manager now.

Access to fft plan cacher singleton has to go through
device manager now.

The driver API in CUDA is not thread safe(or we are using it incorrectly). We
need to protect these calls with mutexes. The cuSparse library also needs
to be protected with these locks.

Implemented in all currently available backends.

af::array::scalar<T>() is changed to use af_get_scalar instead
of fetching entire array and the copying over only first element.

The IN name is defined in the minwindef.h header. Transform also defines
an IN variable which conflicts with this macro and causing compilation
errors.

* Convert CUDA JIT to use nvrtc instead of nvvm

* Fixing underflow/overflow issues with complex math in opencl jit

* Style changes for JIT in CUDA and OpenCL backends

* Use std::string instead of char * for storing text of the jit.cuh file

nvrtc libs are not being found correctly on Ubuntu 16.04. Repurposed
the old FindNVVM.cmake file to locate nvrtc libs.

Differentiate between forge version and tag due to issues with OSX installer
scripts.

Included NIH as sponsor.