Add RMSNorm StableHLO converter (#479)

[michalharakal]

Closes #479.

Summary

Adds a real RMSNorm lowering so modern transformer exports (Llama / Mistral / Qwen / Gemma — every open-weight LLM family uses RMSNorm rather than LayerNorm) stop dropping through the converter registry's "no converter found" path.

rms  = sqrt(mean(x^2, axis) + eps)
out  = scale * x / rms      (scale operand optional)

Emission style matches the softmax fix (#467) and the rest of the emitter: reductions go through `stablehlo.custom_call @reduce_mean`, the reduced tensor is broadcast back to the input shape via `stablehlo.broadcast_in_dim` for the final divide, and the epsilon is materialized as a scalar constant broadcast into the reduced shape.

Two commits

1. Failing test — `RmsNormConverterTest`

• `rmsNorm_operation_is_supported_by_neural_net_converter` — asserts the registry recognizes `rmsNorm`, `rms_norm`, and `RMSNorm`.
• `rmsNorm_with_scale_lowers_to_real_ops` — builds a 2×4 FP32 graph with an RMSNorm node and a per-channel scale operand, asserts the emitted MLIR contains `@reduce_mean`, `sqrt`, `divide`, `broadcast_in_dim`, and `multiply`, and is not labelled `Unsupported operation rmsNorm`.
• `rmsNorm_without_scale_still_normalizes` — the scale operand is optional; dropping it must still produce the core norm.

Red against the pre-fix converter: no handler claims `rmsNorm` at all.

2. The fix

• `NeuralNetOperationsConverter.supportedOperations` adds `rmsNorm`, `rms_norm`, `RMSNorm`, `RmsNorm`.
• New `convertRmsNorm` method handles axis normalization against rank (negative axes; `IntArray normalized_shape` parameter too), uses 1e-6 as the default epsilon (Llama family convention; overridable via `eps` / `epsilon`), and omits the final affine multiply when no scale operand is present so the emitted MLIR stays faithful to the input graph.

Test plan

• `RmsNormConverterTest` — 3/3 green
• Full `:skainet-compile:skainet-compile-hlo:jvmTest` — green, no regressions
• CI: full multiplatform build

Out of scope

• Real `stablehlo.reduce` region bodies. Every reduction in the emitter uses `custom_call @reduce_*` today; migrating all reductions to proper regions is a separate, larger refactor.
• Fusing RMSNorm with the following matmul / attention. That's an IREE-side optimization.
• Quantized RMSNorm — depends on further P0-1 track work.

🤖 Generated with Claude Code