magma: add -Xfatbin -compress-all flags when compiling with cuda

[Markus92]

When compiling magma with a lot of cuda architectures (we got 6 in our heterogenous cluster!) the library can get over 2GB of code causing linker errors:

lib/libmagma.so: PC-relative offset overflow in PLT entry for `_Z24herk_template_batched_tnI6float2Li8ELi12ELi24ELi24ELi8ELi1ELi8ELi12ELi8ELi12ELi1ELi0EEv12magma_uplo_tiiPKPKT_iiiS6_iiiPPS2_iiiS2_S2_iP11magma_queue'
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/magma.dir/build.make:23918: lib/libmagma.so] Error 1
make[2]: Leaving directory '/tmp/spackadmin/spack-stage/spack-stage-magma-2.9.0

The people over at Nix had the same issue and looked at how PyTorch solved it, as PyTorch ships with a dozen architectures for their conda builds.

This PR adds these flags for the CUDA builds and this fixed my linker issues.