[llvm] [AMDGPU] Enable volatile and non-temporal for loads to LDS (PR #153244)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

We're now confined to the DMAs, and I'm pretty sure the cache bits do what we want them to do, no? 

This is coming from the fact that empirically, setting such cache bits is needed for performance, and I'm trying to avoid a big architecture-specific switch-case on `Aux`

https://github.com/llvm/llvm-project/pull/153244