Fix cuTensor, cuTensorNet and cuStateVec when using local Toolkit

[bjoe2k4]

Currently when trying to use cuTENSOR with a local version of the CUDA toolkit, this fails with:

ERROR: LoadError: InitError: UndefVarError: `CUDA_Runtime` not defined

due to #2058 being incomplete.

Versions used:

• Julia 1.10.1
• Cuda Toolkit v12.3.2 loaded via Lmod on a HPC cluster

There are some more occurances of CUDA_Runtime in CUDA.jl, but I haven't had any errors due to these: https://github.com/search?q=repo%3AJuliaGPU%2FCUDA.jl%20CUDA_Runtime.&type=code

I have tested this PR by using cuTENSOR (also locally provided as Lmod module), and fixed cuTensorNet (currently not compatible with cuTensor v2?) and cuStateVec due to similarity without testing.

[maleadt]

Thanks. The original idea was to conditionally define CUDA_Runtime, as is done in CUDA.jl itself, but apparently I forgot about that in the subpackages.

[maleadt]

cuTensorNet (currently not compatible with cuTensor v2?)

That's correct; there's been no release (by NVIDIA) yet of a cuQuantum compatible with CUTENSOR v2. The JLLs should encode those compatibility constraints, but in your case you'll have to manage that yourself.