`xsmm.fused_brgemm` lacks input validation

[adam-smnk]

The snippet taken from xsmm-quarternary-bf16.mlir test:

func.func @entry(%arg0: memref<64x4x4xbf16>, %arg1: memref<64x2x4x2xbf16>, %arg2: memref<4xbf16>, %arg3: memref<4x4xbf16>) {
  %c16_i64 = arith.constant 16 : i64
  %func = xsmm.fused_brgemm.dispatch [4, 4, 4, 4, 4, 4][add, relu]
    flags = (vnni_b) binary_flags = (bcast_col_in0) unary_flags = (none) data_type = bf16
  xsmm.fused_brgemm(data_type = bf16, %func, %arg0, %arg1, %arg2, %arg3, %c16_i64) : (i64, memref<64x4x4xbf16>, memref<64x2x4x2xbf16>, memref<4xbf16>, memref<4x4xbf16>, i64) -> ()
  return
}

The output of xsmm.fused_brgemm operation is stored in %arg2 which is a 1D tensor instead of 2D. Probably XSMM ops validation is too relaxed.