Finalize GPU kernels early with -fno-gpu-rdc

- finalize GPU kernels at compile-time with -fno-gpu-rdc
- abolish using hipcc as linker
- use gcc as linker, link with HIP/HCC runtime
- do NOT use --gc-sections at link time to prevese .kernel section for GPU kernels

- amend missing bazel logic between downstream ROCm fork and TF mainline