Don't hardcode path to libcudart

[Infinoid]

This patch fixes cuda builds on Debian bullseye (testing) with the nvidia-cuda-toolkit and libcudart10.2 packages installed.

Cmake sees /usr/bin/nvcc and decides that cuda is installed in /usr. Taco then assumes libcudart.so is in /usr/lib64/, but Debian puts it in /usr/lib/x86_64-linux-gnu/, instead. So the build fails with the following errors:

make[2]: *** No rule to make target '/usr/lib64/libcudart.so', needed by 'bin/taco'.  Stop.
make[2]: *** No rule to make target '/usr/lib64/libcudart.so', needed by 'bin/taco-tensor_times_vector'.  Stop.
make[2]: *** No rule to make target '/usr/lib64/libcudart.so', needed by 'bin/taco-test'.  Stop.

Cmake sets CUDA_LIBRARIES to the cuda RT library, so we can use that instead of hardcoding paths. With this, the libtaco.so library links fine.
However, the executables (bin/taco, bin/taco-test, etc) failed to build with the following error:

/usr/bin/ld: ../bin/taco: hidden symbol `cudaFree' in /usr/lib/gcc/x86_64-linux-gnu/10/../../../x86_64-linux-gnu/libcudart_static.a(libcudart_static.a.o) is referenced by DSO

It seems libtaco.so is using symbols from libcudart.so without linking against it, but the executable uses libcudart_static.a, and they don't mix.

I was able to fix this by having libtaco.so link to libcudart.so directly, by changing INTERFACE to PUBLIC. In this case, libtaco.so pulls in libcudart.so directly, and the applications get it for free. Since libtaco.so depends on libcudart.so anyway, I don't think there's a downside to this.

With this, it works properly for me with both nvidia's "cuda*.run" installation script (/usr/local/cuda/lib64) and debian's cuda packages (/usr/lib/x86_64-linux-gnu).

[Infinoid]

I've only tested this on very recent Linux distros. It would be good to test this on older linux (ubuntu 18.04 or earlier), and other OSes too.

[Infinoid]

I tested this on ubuntu 18.04, it works fine.

On debian testing with debian-packaged cuda:
-- Found CUDA: /usr/lib (found version "11.1")

On debian testing with nvidia .run file:
-- Found CUDA: /usr/local/cuda (found version "11.2")

On ubuntu 18.04 with nvidia .run file:
-- Found CUDA: /usr/local/cuda (found version "11.2")

and the build completed successfully in all 3 cases.

I noticed two side effects of the patch:

• PATH, LD_LIBRARY_PATH and LIBRARY_PATH don't need to point at /usr/local/cuda to build taco any more, the cmake path detection is enough. (but they are still needed to run tests)
• It seems to be linking with libcudart.a, rather than libcudart.so

Is the static linking a problem?