fix: use out kwarg of _scaled_mm

`C = torch._scaled_mm(A, B, scale, scale)` never triggers the actual matrix multiplication in Gaudi’s lazy execution mode, since operations are deferred until required (e.g., when the output is evaluated, or when an H↔D sync is needed). To ensure the matrix multiplication is executed properly, the code should be written as `torch._scaled_mm(A, B, scale, scale, out=C)`. This change might affect some of the previous benchmarks, but it also aligns with how other dtypes are measured. This PR addresses the FP8 benchmark issue described in https://github.com/stas00/ml-engineering/issues/115