disable tensor op matmul kernels when input and output tensors are in f32 data type to avoid potential accuracy loss

GitOrigin-RevId: 36859cba5a526a7778e12c03ac32815144fe0505