Drop __fd__ , __fhd__ etc. in favor of something more palletable

[eyalroz]

My shorthands for __forceinline__, __device__ and __host__ are quite convenient, but perhaps it's better I don't force them on people - even at the price of making my code more verbose.

I'd appreciate input from whoever visits this page.

[eyalroz]

I'll switch to the approach used in many other libraries:

• No more defining-then-undefining execution space specifiers; only defining them once, and keeping them defined.
• Avoiding conflicts with other libraries/code's defines, by using our library name's prefix: KAT_WHATEVER.
• No more double-prefixing-underscore, that's illegitimate.
• Not using the original specifiers, since those will make compilation fail with non-CUDA compilers - for those files usable with regular C++ compilers.

So, we'll have KAT_HD, KAT_FHD, KAT_DEV, KAT_HOST, KAT_FD and so on. It's not super-readable, but at least people will know it's a library-specific decoration.