[llvm] [AMDGPU] Let LDSDMA ops cross a sched_barrier when the mask allows it (PR #203255)

[Arseniy Obolenskiy via llvm-commits]

aobolensk wrote:

> If we're adding new mask entries - or adding a special mask category for DMAs - we should make sure to update the MLIR wrappers as well

Updated mask docs in both MLIR and llvm/docs/AMDGPUUsage.rst

https://github.com/llvm/llvm-project/pull/203255