feat: protoype for dynamically sized arrays

[avik-pal]

module @reactant_sum attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<?x2xf32> {enzymexla.memory_effects = [], tf.aliasing_output = 1 : i32}) -> (tensor<f32>, tensor<?x2xf32>) attributes {enzymexla.memory_effects = []} {
    %cst = stablehlo.constant dense<0.000000e+00> : tensor<f32>
    %0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<?x2xf32>) -> tensor<2x?xf32>
    %1 = stablehlo.reduce(%0 init: %cst) applies stablehlo.add across dimensions = [0, 1] : (tensor<2x?xf32>, tensor<f32>) -> tensor<f32>
    %2 = stablehlo.transpose %0, dims = [1, 0] : (tensor<2x?xf32>) -> tensor<?x2xf32>
    return %1, %2 : tensor<f32>, tensor<?x2xf32>
  }
}

[avik-pal]

generally most ops wont work with unbounded dynamism. One thing we can do is to store the trace and on input run shape refinement.