feat(hlo): add GroupNorm StableHLO converter

[michalharakal]

Why

groupNorm had no converter — a groupNorm node fell through NeuralNetOperationsConverter to the "Operation not supported" path and never lowered to StableHLO. It's one of the gaps tracked in #751 (left over after #666 closed).

What

Add convertGroupNorm mirroring the existing LayerNorm/RMSNorm decomposition (no custom_call stub):

xg  = reshape(x, [N, G, M])                  # M = (C/G) * prod(spatial)
out = (xg - mean(xg)) / sqrt(var(xg) + eps)  # reduce over the trailing axis
out = reshape(out, [N, C, *spatial]) * scale + offset   # affine optional

• Per-group mean/var reuse the @reduce_mean / @reduce_variance custom_calls (exactly as LayerNorm does) by collapsing each group's channels + spatial into one trailing axis.
• scale/offset (shape C) broadcast over the channel dim (dims=[1]) only.
• Registered the op names (groupNorm/groupNormalization/GroupNormalization/group_norm) in supportedOperations and the dispatch when.

Emitted MLIR for (N=1,C=4,H=2,W=2), num_groups=2:

%v0  = stablehlo.reshape %arg0 : (tensor<1x4x2x2xf32>) -> tensor<1x2x8xf32>
%v1  = stablehlo.custom_call @reduce_mean(%v0) {dimensions = [2], keepdim = false} : tensor<1x2xf32>
... subtract / reduce_variance / +eps / sqrt / divide ...
%v11 = stablehlo.reshape %v10 : (tensor<1x2x8xf32>) -> tensor<1x4x2x2xf32>
%v12 = stablehlo.broadcast_in_dim %arg1, dims = [1] : (tensor<4xf32>) -> tensor<1x4x2x2xf32>   # scale
%v13 = stablehlo.multiply %v11, %v12 : tensor<1x4x2x2xf32>
%v14 = stablehlo.broadcast_in_dim %arg2, dims = [1] : (tensor<4xf32>) -> tensor<1x4x2x2xf32>   # offset
%v15 = stablehlo.add %v13, %v14 : tensor<1x4x2x2xf32>

Tests

Adds GroupNormConverterTest (structural, mirrors LayerNormConverterTest): asserts the real reshape + reductions + broadcast + sqrt + divide + affine, and that it never hits the stub / unsupported path. The full skainet-compile-hlo jvmTest suite passes locally on develop.

Notes / follow-ups (per #751)

• Numerical (IREE execution) validation will come via the downstream skainet-iree-conformance harness.
• upsample/interpolate and RoPE remain — tracked in StableHLO converter: missing ops RoPE, groupNorm, upsample/interpolate (gaps remaining after #666) #751.

Closes part of #751.