[Performance] Estimate the cost of Atomic ops with the MMA layout.

[LiyangLingIntel]

To address the comment #2312 (comment).

To fix the issue that in certain gemm cases, ConvertLayoutOp requires oversized shared local memory. In #2312, we use Atomic with mma layout to eliminate ConvertLayoutOp.

In this issue, we need to estimate the cost of Atomic ops with the MMA layout, compare it with the cost of ConvertLayout + AtomicRMW. With the knowledge about the performance difference, we can decide if a cost module needed or not in that pass.