Removal of the previous kernel registration scheme unearthed mistakes in
the new one, which were:
- The new kernel registration code relied on the binder expansion phase,
  which didn't happen because the registration code was already
  generated by the binder.
- The kernel handle passed to CUDA_Register_Function was the first eight
  bytes of the code of the host-side procedure representing the kernel
  rather than its address.

Tested on x86_64-pc-linux-gnu, committed on trunk

gcc/ada/

	* bindgen.adb (Gen_CUDA_Init): Remove code generating CUDA
	definitions.
	(Gen_CUDA_Defs): New function, generating definitions
	initialized by Gen_CUDA_Init.
	(Gen_Output_File_Ada): Call Gen_CUDA_Defs instead of
	Gen_CUDA_Init.
	(Gen_Adainit): Call Gen_CUDA_Init.