[llvm] [AMDGPU] 4-align TTMP triples (PR #132759)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Aligning them on a 4-TTMP boundary matches what we do for SGPRs

We probably should have the unaligned SGPR classes

https://github.com/llvm/llvm-project/pull/132759