[Umbrella] CIR CUDA Support

[AdUhTkJm]

Currently we can compile some real CUDA programs. However, CUDA support is not fully complete, and there are still several things to do:

• CallConvLowering is needed for both x86 and NVPTX. x86 support is incomplete and NVPTX is fully NYI (we can currently get past this if we don't pass any structs around by value.)
• Global variable registration hasn't been done.
• We need to support more corner cases and flags, such as -fgpu-rdc and -foffload-via-llvm.
• We need to emit NVVM metadata.