reduce of broadcast

from neural gcm

```
%2003 = stablehlo.broadcast_in_dim %2002, dims = [1, 2, 3] : (tensor<128x2x1xf32>) -> tensor<128x128x2x1xf32>
%2004 = stablehlo.reshape %2003 : (tensor<128x128x2x1xf32>) -> tensor<128x256x1xf32>
%2005 = stablehlo.reduce(%2004 init: %cst_304) applies stablehlo.add across dimensions = [1] : (tensor<128x256x1xf32>, tensor) -> tensor<128x1xf32>
```

we should do the reduce first and then broadcast. this should be done in 2 parts:

1. fuse reshape reduce

```
%2004 = stablehlo.reshape %2003 : (tensor<128x128x2x1xf32>) -> tensor<128x256x1xf32>
%2005 = stablehlo.reduce(%2004 init: %cst_304) applies stablehlo.add across dimensions = [1] : (tensor<128x256x1xf32>, tensor) -> tensor<128x1xf32>
```

should become `reduce(%2003) dims = [1, 2]`

2. reduce(bcast) where all the reduction dims are part of the original tensor, we should reduce and then bcast

Start from the StableHLO broadcast_in_dim, reshape, and reduce sequence in the issue and locate the rewrite entry point that handles these operations. Implement the two requested transformations: fuse reshape and reduce dimensions, then reduce before broadcasting when reduction dimensions come from the original tensor. Done means the resulting form performs the reduction before the broadcast.

Written by the indexing model from the issue text.