Support passing structs as kernel `launch()` arguments

[leofang]

xref:

cuda-python/cuda_core/cuda/core/_kernel_arg_handler.pyx

Line 356 in 060578a

# TODO: support ctypes/numpy struct

Try NumPy first, then ctypes.

Note: This is for passing single structs by value. For passing an array of structs by pointer (to gmem) it's a separate task.