Batched AD crash for a simple MLP

```mlir
module @reactant_batched... attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
func.func private @neuralnetwork(%arg0: tensor<2x4xf32> {enzymexla.memory_effects = []}, %arg1: tensor<4xf32> {enzymexla.memory_effects = []}, %arg2: tensor<4x3xf32> {enzymexla.memory_effects = []}, %arg3: tensor<3xf32> {enzymexla.memory_effects = []}, %arg4: tensor<4x2xf32> {enzymexla.memory_effects = []}) -> (tensor<4x3xf32>, tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>, tensor<4x2xf32>) attributes {enzymexla.memory_effects = []} {
%0 = stablehlo.dot_general %arg0, %arg4, contracting_dims = [0] x [1], precision = [DEFAULT, DEFAULT] : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<4x4xf32>
%1 = stablehlo.broadcast_in_dim %arg1, dims = [0] : (tensor<4xf32>) -> tensor<4x4xf32>
%2 = stablehlo.add %0, %1 : tensor<4x4xf32>
%3 = stablehlo.tanh %2 : tensor<4x4xf32>
%4 = stablehlo.broadcast_in_dim %arg3, dims = [0] : (tensor<3xf32>) -> tensor<3x4xf32>
%5 = stablehlo.dot_general %arg2, %3, contracting_dims = [0] x [0], precision = [DEFAULT, DEFAULT] : (tensor<4x3xf32>, tensor<4x4xf32>) -> tensor<3x4xf32>
%6 = stablehlo.add %5, %4 : tensor<3x4xf32>
%7 = stablehlo.transpose %6, dims = [1, 0] : (tensor<3x4xf32>) -> tensor<4x3xf32>
return %7, %arg0, %arg1, %arg2, %arg3, %arg4 : tensor<4x3xf32>, tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>, tensor<4x2xf32>
}
func.func @main(%arg0: tensor<2x4xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"], tf.aliasing_output = 1 : i32}, %arg1: tensor<4xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"], tf.aliasing_output = 2 : i32}, %arg2: tensor<4x3xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"], tf.aliasing_output = 3 : i32}, %arg3: tensor<3xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"], tf.aliasing_output = 4 : i32}, %arg4: tensor<4x2xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"]}) -> (tensor<4x2x3xf32>, tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>) attributes {enzymexla.memory_effects = ["read", "write", "allocate", "free"]} {
%cst = stablehlo.constant dense<[[[1.000000e+00, 0.000000e+00], [1.000000e+00, 0.000000e+00], [1.000000e+00, 0.000000e+00], [1.000000e+00, 0.000000e+00]], [[0.000000e+00, 1.000000e+00], [0.000000e+00, 1.000000e+00], [0.000000e+00, 1.000000e+00], [0.000000e+00, 1.000000e+00]]]> : tensor<2x4x2xf32>
%0:5 = enzyme.fwddiff @neuralnetwork(%arg0, %arg1, %arg2, %arg3, %arg4, %cst) {activity = [#enzyme, #enzyme, #enzyme, #enzyme, #enzyme], ret_activity = [#enzyme, #enzyme, #enzyme, #enzyme, #enzyme, #enzyme], width = 2 : i64} : (tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>, tensor<4x2xf32>, tensor<2x4x2xf32>) -> (tensor<2x4x3xf32>, tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>)
%1 = stablehlo.slice %0#0 [1:2, 0:4, 0:3] : (tensor<2x4x3xf32>) -> tensor<1x4x3xf32>
%2 = stablehlo.transpose %1, dims = [2, 1, 0] : (tensor<1x4x3xf32>) -> tensor<3x4x1xf32>
%3 = stablehlo.slice %0#0 [0:1, 0:4, 0:3] : (tensor<2x4x3xf32>) -> tensor<1x4x3xf32>
%4 = stablehlo.transpose %3, dims = [2, 1, 0] : (tensor<1x4x3xf32>) -> tensor<3x4x1xf32>
%5 = stablehlo.reshape %4 : (tensor<3x4x1xf32>) -> tensor<3x1x4xf32>
%6 = stablehlo.reshape %2 : (tensor<3x4x1xf32>) -> tensor<3x1x4xf32>
%7 = stablehlo.concatenate %5, %6, dim = 1 : (tensor<3x1x4xf32>, tensor<3x1x4xf32>) -> tensor<3x2x4xf32>
%8 = stablehlo.transpose %7, dims = [2, 1, 0] : (tensor<3x2x4xf32>) -> tensor<4x2x3xf32>
return %8, %0#1, %0#2, %0#3, %0#4 : tensor<4x2x3xf32>, tensor<2x4xf32>, tensor<4xf32>, tensor<4x3xf32>, tensor<3xf32>
}
}
```

Use the standalone MLIR module in the issue as the reproducer and determine which Enzyme-JAX automatic-differentiation pipeline stage crashes during the batched fwddiff call. Compare the batched MLP case with a non-batched case before tracing the responsible compiler subsystem. Done means the supplied reproducer completes without crashing and its derivative result is validated.

Written by the indexing model from the issue text.