RMSNorm and SwiGLU (Llama's MLP Block)

Llama, Mistral and most modern LLMs swapped two classic pieces for cheaper or better ones:

• RMSNorm replaces LayerNorm: no mean subtraction, no bias. $\text{RMSNorm}(x) = \frac{x}{\sqrt{\text{mean}(x^2) + \epsilon}} \odot g$, computed over the last axis.
• SwiGLU replaces the ReLU/GELU MLP: $\text{SwiGLU}(x) = \big(\text{SiLU}(xW_1) \odot xW_3\big) W_2$, where $\text{SiLU}(z) = z \cdot \sigma(z)$.

Implement silu(z), rms_norm(x, g, eps=1e-6) and swiglu_mlp(x, W1, W3, W2), then llama_block_mlp(x, g, W1, W3, W2), the residual sub-block $x + \text{SwiGLU}(\text{RMSNorm}(x))$.