convolution with batch_group_count -> regular convolution

If the kernel is repeated along the kernel_output_dim, then we can drop the batch group count and permute the output to get the desired output

```
module {
func.func @main(%arg0: tensor<6x3x2x5x4x4xf32>) -> tensor<6x3x7x5x2x2xf32> {
%cst = stablehlo.constant dense<3.000000e+00> : tensor<3x3x2x210xf32>
%0 = stablehlo.transpose %arg0, dims = [5, 4, 2, 1, 3, 0] : (tensor<6x3x2x5x4x4xf32>) -> tensor<4x4x2x3x5x6xf32>
%1 = stablehlo.reshape %0 : (tensor<4x4x2x3x5x6xf32>) -> tensor<4x4x2x90xf32>
%2 = stablehlo.convolution(%1, %cst) dim_numbers = [0, 1, f, b]x[0, 1, i, o]->[0, 1, f, b], window = {stride = [1, 1], pad = [[0, 0], [0, 0]], rhs_dilate = [1, 1]} {batch_group_count = 30 : i64, feature_group_count = 1 : i64, precision_config = [#stablehlo, #stablehlo]} : (tensor<4x4x2x90xf32>, tensor<3x3x2x210xf32>) -> tensor<2x2x210x3xf32>
%3 = stablehlo.reshape %2 : (tensor<2x2x210x3xf32>) -> tensor<2x2x7x5x6x3xf32>
%4 = stablehlo.transpose %3, dims = [4, 5, 2, 3, 1, 0] : (tensor<2x2x7x5x6x3xf32>) -> tensor<6x3x7x5x2x2xf32>
return %4 : tensor<6x3x7x5x2x2xf32>
}
}
```

No source files, entry points, or tests are named. Start by locating the compiler handling for StableHLO convolution operations with batch_group_count, then use the supplied MLIR module as the reproducer; done means the equivalent regular-convolution form preserves the shown output shape and semantics.

Written by the indexing model from the issue text.