Pass GPU_CFLAGS to CUDA linker

[BCSharp]

Currently GPU_CFLAGS (containing --gpu-code and --gpu-architecture) are being passed to nvcc during compilation, but not when producing the shared object (.so) library. As a result, nvcc's defaults are being used, which may or may not match what the object (.o) files may contain, depending on the CUDA version used. In case of a mismatch, warnings are generated.

A safe way to avoid the mismatch is to always link with the same GPU_CPLAGS as used during compilation.

Example warnings on CUDA-11.2 when GPU_ARCHS:=sm_61 and GPU_PTX_ARCH:=compute_61:

nvcc -Xcompiler "-fPIC" --shared --output-file release/libcuda-crypt.so release/crypt-dlink.o release/verify.o release/poh_verify.o release/gpu_ctx.o release/sign.o release/seed.o release/keypair.o
nvlink warning : SM Arch ('sm_52') not found in 'release/verify.o'
nvlink warning : SM Arch ('sm_52') not found in 'release/poh_verify.o'
nvlink warning : SM Arch ('sm_52') not found in 'release/gpu_ctx.o'
nvlink warning : SM Arch ('sm_52') not found in 'release/sign.o'
nvlink warning : SM Arch ('sm_52') not found in 'release/seed.o'
nvlink warning : SM Arch ('sm_52') not found in 'release/keypair.o'

This PR adds $(GPU_CFLAGS) to nvcc parameters used in the linking phase.

[ryoqun]

@sakridge could you review this? this looks nice improvement to the build system?