Register silu alias in ActivationOperationsConverter (P1)

[michalharakal]

Context

`ActivationOperationsConverter` currently registers the activation function as `swish` and lowers it to `x * sigmoid(x)`. Every Llama / Mistral / Qwen / Gemma family model names the same function `silu` (SiLU — Sigmoid Linear Unit) in their graph metadata, not `swish`. PyTorch's `F.silu` and HuggingFace's `ACT2FN["silu"]` both emit `silu`.

Today a traced Llama model with `silu` activation falls through the converter registry's "no converter found" path.

This PR

Strictly additive, scoped to one line of code plus one line of test:

1. Add `"silu"` and `"SiLU"` to `ActivationOperationsConverter.supportedOperations`.
2. Add the `"silu"` case to the `when` in `convert`, dispatching to the existing `convertSwish` (same lowering: `x * sigmoid(x)`).
3. Extend `ActivationOperationsConverterTest`: assert `registry.isSupported("silu")` and that a graph using `silu` emits a module containing both `stablehlo.exponential` and `stablehlo.multiply` — the same assertions as the existing Swish test, just with the alias name.

Why this is its own issue

Scoping to a single alias keeps the change mechanically reviewable in under 30 seconds and keeps any follow-on work (proper GELU/gelu_tanh registration, unified activation metadata table) unblocked by this trivial PR.

Out of scope

• New activation functions.
• Any other alias (gelu_new, relu6, etc.). Handle in follow-ups if traced models surface them.