cudnn: add rpath to helper libraries for cudnn 8

[Markus92]

CuDNN versions prior to 7 shipped with one library: libcudnn.so. This made linking easy as there is only one thing to link to.

CuDNN 8 changed this: the CuDNN library was split into 6 libraries and a shim libcudnn.so. To make sure libcudnn.so can find it's helper libraries, it has its rpath set to $ORIGIN. As all libs come into the same folder, this works fine for the most part. However, the individual split libraries do not have an rpath set, so they cannot see each other. When using pytorch this can lead to library not found errors during runtime, as the libraries do link to each other: e.g. libcudnn_adv_infer.so links to libcudnn_ops_infer.so.8

CuDNN 9 has the rpath set to $ORIGIN in all helper libraries off-the-shelf, so is not affected by this behavior. This PR adds that behavior to CuDNN 8.

[Markus92]

Thanks for the comments, @haampie ! Took care of them hopefully

[haampie]

Would be good to test drive a bit, cause patchelf's mechanism for growing the dynamic section is a rather involved operation. Plus look if the file size is much larger than before (probably won't matter cause of how big the rest of cuda is).

I don't have an nvidia gpu to test on though

[Markus92]

File sizes differ a bit but not much, 7mb on a 2.3gb package for CuDNN 8.9.7

Some simple tests seem to indicate it works (simple MNIST example). I'll deploy it across our users and report back if it fixes their issues.

[haampie]

sounds good to me