Implement missing host-side math functions

[pvelesko]

Implement missing host math funcs

* Compiling the bitode library to host code leads to linking errors
some of the OpenCL natives being used in the library code are not
available for the CPU, resulting in linking errors

* OCML does not meant to be compiled in C++ mode
for this, we create a drop-in opencl header defining the missing
datatypes

* Since OCML uses headers in the same directory as the source files,
we must copy the necessary source files into a new directory so that
our replacement headers get picked up instead of the originals

* Implement a new header for mapping math funcs to host-side ocml

* Discover an additional way to trigger 128 bit error

* Discover a discrepancy in host and device func results