[build] cuda: Introduce CUDA_SRCS macro utilizing parallelized device code compilation

Instead of a single graph node launching NVCC to compile .cu for both host and all device architectures
CUDA_SRCS generates multiple nodes:
- node per each device architecture producing PTX and CUBIN
- node merging all PTX and CUBIN files into a single FATBIN blob
- node producing .cpp with host code
- node compiling host .cpp with embedded FATBIN blob

CUDA_ARCHITECTURES variable is used to determine the list of architectures to compile device code for.

ISSUE:
commit_hash:0a4c2797dd238ae062482af30694df6978301278

0 files changed, 0 insertions, 0 deletions