Remove more GPU/CUDA/ROCm attribute guards from xla/service/gpu

[copybara-service[bot]]

Remove more GPU/CUDA/ROCm attribute guards from xla/service/gpu

• This removes if_gpu_is_configured guards from targets that are only supposed to be built for GPU. (Also tags them as gpu so that they get automatically excluded from the CPU build)
• This removes a bunch of local_defines = if_cuda_is_configured([GOOGLE_CUDA]) parameters from targets that don't use the GOOGLE_CUDA preprocessor definition (anymore).
• Same for TENSORFLOW_USE_ROCM
• It tags many gpu-only targets with gpu so that they won't be built in a CPU-only build.