Fix RCCL install, linear.py logic, CMake custom extension, update requirement for FP8 compute

[mawong-amd]

This small PR does the following:

1. Fix buggy RCCL installation in Dockerfile: when installed as a package, it has to be installed twice.
2. Update linear.py, removing duplicated custom kernel invocation logic which has been moved inside tuned_gemm.py. It also generalizes the use of tuned_gemm.py (and hence tuning/custom kernel invocation) in place of a direct call to torch.nn.functional.Linear when bias is not fused.
3. Correct the CMake logic for the custom extension _custom_C so it is not built on CUDA.
4. Add pandas as a ROCm-specific requirement due to its use in FP8 linear methods, specifically when they relate to tuning.