[FEA]: Factor the generated CUDA shim into an ABI mode shared library #3253

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

cuda.bindings

Is your feature request related to a problem? Please describe.

I'd like to be able to repeat the cuda.h free, cuGetProcAddress approach of accessing the driver API in another library.

Describe the solution you'd like

A header-free C layer that dlopens libcuda.so.1, resolves entry points through cuGetProcAddress, caches the function pointers, and exposes ABI-reduced C symbols. This is a pattern here and other places and should have a single source of truth.

Today this is behind private cdef functions inside a Python extension rather than a C ABI another shared library or JIT can call.

Describe alternatives you've considered

Rolling another copy again, which seems suboptimal.

Additional context

No response