[P1] Enhance TOSA constant folding effectivenes

[guosran]

Hi all :) During the implementation of the TOSA to Taskflow pipeline in #245 , I have identified the standard tosa-layerwise-constant-fold pass does not currently fold simple constant subexpressions when using arith.constant as operands.

• Input TOSA IR:

func.func @const_fold_test() -> tensor<4xf32> {
  %cst1 = arith.constant dense<[1.0, 2.0, 3.0, 4.0]> : tensor<4xf32>
  %cst2 = arith.constant dense<[10.0, 20.0, 30.0, 40.0]> : tensor<4xf32>
  
  // This addition should be folded!
  %folded = tosa.add %cst1, %cst2 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  return %folded : tensor<4xf32>
}

• Current Suboptimal Output:

// Actual result from current pipeline: runtime calculation
affine.for %arg1 = 0 to 4 {
  %2 = affine.load %0[%arg1] : memref<4xf32>
  %3 = affine.load %1[%arg1] : memref<4xf32>
  %4 = arith.addf %2, %3 : f32   // <--- Suboptimal: Runtime addition
  affine.store %4, %alloc[%arg1] : memref<4xf32>
}

• Expected Target Output:

// Desired Result: Pure constant propagation
memref.global "private" constant @__constant_sum : memref<4xf32> = dense<[11.0, 12.0, 13.0, 14.0]>
func.func @const_fold_test(%arg0: memref<4xf32>) {
  %0 = memref.get_global @__constant_sum : memref<4xf32>
  memref.copy %0, %arg0 : memref<4xf32> to memref<4xf32>
  return
}

Which indicates the current solution is not optimal.