Simplest CUDA half type conflict resolution implementation.

[oxt3479]

This implementation will simply place half into a namespace if compiling for the cuda kernel to prevent conflicts with the identically named cuda half type.

Note: A few definitions needed to be moved around to avoid placing functions within the namespace and breaking existing testing. (operator <<, >>, and printBits)

Issues moving forward:

1: Testing of the cuda::half type's behaviour can be added, however the cuda half type is more picky about comparison and type conversion. For instance:

half h1;
...
assert (h1 == 1);

Will produce compiler errors if written with the cuda half (but not for imath::half). It must instead be written this way:

half h1;
...
assert(float(h1) == 1);

2: The cuda half type also can not be created on host code and therefore printing using cout will not be trivial. Most tests use this heavily.

The bottom line: making the cuda::half behave just as the imath::half type requires considerable reworking and does not seem to be fully attainable for many reasons.
Certainly testing can be done it will just not be the same as the testing already implemented under HalfTest/

[lgritz]

indentation