Build CUDA intermediate files with a reproducible name

[fwyzard]

By default, intermediate files built by nvcc -dc and nvcc -dlink are stored in a temporary directory and have a randomly generated name, e.g.

/tmp/tmpxft_00031a01_00000000-6_kernels.cpp1.ii

These file names are then used to generate unique symbols names, e.g.

T __cudaRegisterLinkedBinary_42_tmpxft_00031a01_00000000_6_kernels_cpp1_ii_478d4aee

and so end up influencing the final binaries.

Experimenting with nvcc, it seems that this can be avoided using the -keep -keep-dir $TEMP_DIR options: in this case the temporary names are relative to the TEMP_DIR directory, and so depend only on the compilation target name, and should remain unchanged across recompilation.

Depending how scram invokes nvcc, a possibility could be to pass the options -keep -keep-dir $CMSSW_BASE/tmp/$SCRAM_ARCH (or a dedicated subdirectory).