[llvm] [AMDGPU] Skip the iterative atomic scan on native LDS atomics (PR #227222)

[Dmitry Sidorov via llvm-commits]

MrSidims wrote:

The fix comes from an observation, that few HeCBenches are running faster with atomic-optimizer-strategy=None than default.
Example on gfx90a:
threadfence-hip: None on current main is 1.88x faster
aop-hip: None on current main is 1.66x faster

Path fixes this.

https://github.com/llvm/llvm-project/pull/227222