Support for simple TFLite dynamic shape computations

Request description

In some simple cases we can be left with what appears to be a dynamic shape computation. In these cases it is technically expressible under tosa however requires multiple operations to be folded together and statically analyzed. A simple example is shown below.

func.func @main(%arg0: tensor<?x16xf32>) -> tensor<?x?xf32> {
  %cst_0 = arith.constant dense<0> : tensor<i32>
  %cst_2 = arith.constant dense<2> : tensor<i32>
  %0 = "tfl.shape"(%arg0) : (tensor<?x16xf32>) -> tensor<2xi32>
  %1:2 = "tfl.split"(%cst_0, %0) { num_splits = 2 : i32 } : (tensor<i32>, tensor<2xi32>) -> (tensor<1xi32>, tensor<1xi32>)
  %2 = arith.constant dense<> : tensor<0xi32>
  %3 = "tfl.reshape"(%1#0, %2) : (tensor<1xi32>, tensor<0xi32>) -> tensor<*xi32>
  %4 = tfl.mul(%3, %cst_2) {fused_activation_function = "NONE"} : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
  %5 = arith.constant dense<8> : tensor<i32>
  %6 = "tfl.pack"(%4, %5) {axis = 0 : i32, values_count = 2 : i32} : (tensor<*xi32>, tensor<i32>) -> tensor<2xi32>
  %7 = "tfl.reshape"(%arg0, %6) : (tensor<?x16xf32>, tensor<2xi32>) -> tensor<?x?xf32>
  return %7 : tensor<?x?xf32>
}

This case contains some basic reshaping however the output shape of the final reshape is unknown as it is technically a dynamic value. We need a method to propagate shapes forward for these static cases similar to mhlo.

What component(s) does this issue relate to?

No response

Additional context

No response