`PadDotGeneral` is busted

```mlir
// ./bazel-bin/enzymexlamlir-opt --pass-pipeline="any(enzyme-hlo-generate-td{patterns=pad_dot_general<1>(0);pad_dot_general<1>(1)},transform-interpreter,enzyme-hlo-remove-transform)" test/lit_tests/dus_to_dynamic_pad.mlir
module {
func.func @fn(%arg0: tensor {enzymexla.memory_effects = ["read", "write", "allocate", "free"]}) -> (tensor<32x32xf32>, tensor) attributes {enzymexla.memory_effects = ["read", "write", "allocate", "free"]} {
%c = stablehlo.constant dense<3> : tensor<32x32xi64>
%c_0 = stablehlo.constant dense<2> : tensor<32x32xi64>
%cst = stablehlo.constant dense<0.000000e+00> : tensor
%0 = stablehlo.reshape %arg0 : (tensor) -> tensor<1x1xf32>
%1 = stablehlo.pad %0, %cst, low = [2, 3], high = [29, 28], interior = [0, 0] : (tensor<1x1xf32>, tensor) -> tensor<32x32xf32>
%2 = stablehlo.convert %c_0 : (tensor<32x32xi64>) -> tensor<32x32xf32>
%3 = stablehlo.multiply %2, %1 : tensor<32x32xf32>
%4 = stablehlo.convert %c : (tensor<32x32xi64>) -> tensor<32x32xf32>
%5 = stablehlo.subtract %3, %4 : tensor<32x32xf32>
%6 = stablehlo.pad %0, %cst, low = [3, 2], high = [28, 29], interior = [0, 0] : (tensor<1x1xf32>, tensor) -> tensor<32x32xf32>
%7 = stablehlo.dot_general %6, %5, contracting_dims = [0] x [1], precision = [DEFAULT, DEFAULT] : (tensor<32x32xf32>, tensor<32x32xf32>) -> tensor<32x32xf32>
return %7, %arg0 : tensor<32x32xf32>, tensor
}
}
```

```mlir
test/lit_tests/dus_to_dynamic_pad.mlir:14:10: error: contracting dimension sizes must match for lhs/rhs
%7 = stablehlo.dot_general %6, %5, contracting_dims = [0] x [1], precision = [DEFAULT, DEFAULT] : (tensor<32x32xf32>, tensor<32x32xf32>) -> tensor<32x32xf32>
```

Run the bazel-bin/enzymexlamlir-opt command from the issue against test/lit_tests/dus_to_dynamic_pad.mlir and inspect the PadDotGeneral path used by the pad_dot_general patterns. Trace the reported contracting-dimension mismatch at the stablehlo.dot_general operation; done means the reproducer no longer emits that diagnostic and the relevant lit test passes.

Written by the indexing model from the issue text.