cudaFuncGetAttributes does not work but cuFuncGetAttribute does

[n-eiling]

As mentioned in #15

cudaFuncGetAttributes exec fail in server error is cudaErrorInvalidDeviceFunction，if use cuFuncGetAttribute replace is ok

this likely occurs because runtime API functions do not exist on the server side but are translated to driver API cuModule and cuFunction objects. We need to translate these for every runtime API function that references (if any?) modules or functions.

[zaglc]

Hi @n-eiling , in your recent commit f251ebaef94b6e66ba1012064d9d898e98794f5b you modify implement of cudaFuncs. In my experiment running stable diffusion in pytorch version, I implement cuda_func_get_attributes_1_svc by enumerating CUfunction_attribute and invoking cuFuncGetAttribute, which indeed works.

Moreover, the cuda kernel (device function?) is stored in rm_functions, whose addr differs between client and server. I think the following code is needed otherwise the server will encounter segmentation fault

void *new_func = resource_mg_get(&rm_functions, (void*)func);

Besides, I'm using cuda 11.8 on A4500