[llvm] [AMDGPU] Stop combining arbitrary offsets into PAL relocs (PR #80034)

[Jay Foad via llvm-commits]

jayfoad wrote:

Possible future improvements:
- Change PAL to use RELA relocations.
- Failing that, implement some DAG combines to fold the offset into a GlobalAddressSDNode when we know that the resulting offset will fit.

https://github.com/llvm/llvm-project/pull/80034