Use switches (one already existing, one newly added here) to indicate to
the binder that CUDA support code is to be generated for either the
host side or for the device side. Add an invocation of Adainit on the
device side from Adainit on the host side; similarly for Adafinal.

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

gcc/ada/

	* bindgen.adb: When the binder is invoked for the host, it
	declares imported subprograms corresponding to the Adainit and
	Adafinal routines on the device. Declare string constants and
	expression functions for the Ada source names and the link names
	of these routines. Generate these subprogram declarations (and
	accompanying Import pragmas) in Gen_CUDA_Defs. Generate
	CUDA_Execute pragmas to call these subprograms from the host in
	Gen_Adafinal and Gen_CUDA_Init. When the binder is invoked for the
	device, include a CUDA_Global aspect declaration in the
	declarations of Adainit and Adafinal and use the aforementioned
	link names in the Export pragmas generated for those two routines.
	* debug.adb: Update comments about "d_c" and "d_d" switches.
	* opt.ads: Declare new Boolean variable,
	Enable_CUDA_Device_Expansion. This complements the existing
	Enable_CUDA_Expansion variable, which is used to enable host-side
	CUDA expansion. The new variable enables device-side CUDA
	expansion. It is currently never set during compilation; it is
	only set via a binder switch.
	* switch-b.adb
	(scan_debug_switches): Add new use of the "-d_d" binder switch.
	The new switch and the variable Opt.Enabled_CUDA_Device_Expansion
	follow the existing pattern of the "-d_c" switch and the variable
	Opt.Enabled_CUDA_Expansion. Flag error if both "-d_c" and "-d_d"
	are specified.