Add RMSNorm StableHLO converter (P1)

[michalharakal]

Context

Follow-up to #467 (softmax fix). With softmax lowered correctly, the next correctness gap in the StableHLO emitter for modern transformer exports is RMSNorm. Every Llama / Mistral / Qwen / Gemma family model normalizes its activations with RMSNorm, not LayerNorm, and there is no converter for it today. `ActivationOperationsConverter` and `NeuralNetOperationsConverter` have zero entries for `rmsNorm` / `rms_norm` / `RMSNorm`, so a traced Llama model hits the converter registry's "no converter found" path and fails.

RMSNorm is:

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

(No mean-centering, no offset — that's what distinguishes it from LayerNorm.)

This PR

Add an `RMSNormConverter` (or extend `NeuralNetOperationsConverter`, whichever matches the codebase convention best for normalization ops) that:

1. Accepts operation names `rmsNorm`, `rms_norm`, `RMSNorm`, `RmsNorm`.
2. Takes 1 or 2 operands: `(input)` or `(input, scale)`. If `scale` is absent, emits the norm without the final multiply (scale=1.0).
3. Reads parameters: `eps` (default `1e-6`, typical for Llama/Mistral), `normalized_shape` or `axis` (default: last dim).
4. Emits a lowering using the codebase's existing `custom_call @reduce_mean` style (same as `ReductionOperationsConverter` uses for mean / variance) so it's syntactically consistent with the rest of the emitter today:

```mlir
%x_squared = stablehlo.multiply %x, %x : tensor<...>
%mean_sq = stablehlo.custom_call @reduce_mean(%x_squared) {dimensions = [axis], keepdim = true} : tensor<...>
%eps_const = stablehlo.constant dense<1.0e-06> : tensor
%eps_b = stablehlo.broadcast_in_dim %eps_const, dims = [] : (tensor) -> tensor<...>
%var_eps = stablehlo.add %mean_sq, %eps_b : tensor<...>
%rms = stablehlo.sqrt %var_eps : tensor<...>
%normed = stablehlo.divide %x, %rms : tensor<...>
%out = stablehlo.multiply %normed, %scale : tensor<...> // only when scale operand present
```

5. Handles axis normalization against rank (negative axes).
6. Registers the converter in `StableHloConverterFactory` so `registry.isSupported("rmsNorm")` returns true end-to-end.
7. Unit test in the appropriate test file: graph with an RMSNorm node + optional scale operand, assert the emitted module contains all expected ops and *no* placeholder-constant fakes.

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.