broadcast(convert(...)) to convert(broadcast()) based on types

```
%19 = stablehlo.convert %arg11 : (tensor<32xbf16>) -> tensor<32xf32>
%20 = stablehlo.broadcast_in_dim %19, dims = [2] : (tensor<32xf32>) -> tensor<128x12x32xf32>
```

We should broadcast and then convert when increasing precision

Start by locating the code that emits the StableHLO convert and broadcast_in_dim operations shown in the issue. Trace how operand and result types determine their ordering, then verify that increasing precision broadcasts before converting and that the relevant regression coverage passes.

Written by the indexing model from the issue text.