Expose `cuda.core.Kernel.attributes.num_arguments`

[leofang]

This is a handy approach as suggested by @galv since the driver does not expose any queryable APIs for this purpose:

Hello pipul, there is not a direct API, but you can do this https://github.com/pytorch/pytorch/pull/137318/files#diff-d7302d133bb5e0890fc94de9aeea4d9d442555a3b40772c9db10edb5cf36a35cR424-R429

In other words:

for(size_t i = 0; cudaFuncGetParamInfo(func, i, &param_offset, &param_size) != CUDA_ERROR_INVALID_VALUE; i++) {}
// i now contains the total number of arguments to your kernel.

I'm sure you can translate that to python.

Originally posted by @galv in #543 (reply in thread)

[leofang]

@vzhurba01 we'll need the unsupported APIs to be added to the cython/interface layers. This one uses cudaFuncGetParamInfo.

[leofang]

During an offline discussion Sasha suggested instead of exposing it as Kernel.attributes.num_arguments we can just expose it as Kernel.num_arguments. I suppose either way is fine.