[llvm] [AMDGPU] Tweak handling of GlobalAddress operands in SI_PC_ADD_REL_OFFSET (PR #70960)

[Jay Foad via llvm-commits]

jayfoad wrote:

A possible follow up improvement would be to make SI_PC_ADD_REL_OFFSET take a single GlobalAddress operand, and only split it into (one or two) 32-bit parts in expandPostRAPseudo.

https://github.com/llvm/llvm-project/pull/70960