[llvm] [AMDGPU] Add function attribute to control TBUFFER combining (PR #156454)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

@piotrAMD More fundamental question: is the strictness/relaxed-ness something that's enabled per-*compilation* (that is, per-module) or, at the SPIR-V level, can in differ per-*function* ?

If you can't have functions with different strictness requirements in ore program ... both of these should be module flags, and `AllowsMisalignedMemoryAccesses()` should be changed to take a `Module*` (or a `Function*`) as appropriate.

https://github.com/llvm/llvm-project/pull/156454