Add Laguna S 2.1 model support

[avlp12]

Summary

• add the 48-layer Laguna S 2.1 MLX runtime, including heterogeneous attention heads, YARN/default RoPE, per-head softplus gating, correction-bias routing, shared experts, and mixed KV caches
• sanitize 36,096 individual expert tensors into ordered stacked SwitchGLU projections
• exclude routers, norms, correction biases, and other control tensors from quantization

Validation

• 10 Laguna unit tests pass
• independent tiny PyTorch/MLX parity passes for routing, unbiased weights, attention gating, both RoPE variants, expert ordering, full logits, and cached logits
• maximum observed full/cached logit absolute error: 2.98e-8
• Ruff format/check and git diff check pass

Downstream release evidence

• end-to-end release tooling and publication validation: 343 tests passed, plus 161 subtests passed
• public compact release: Laguna-S-2.1-Alis-MLX-Dynamic-3.306814bpw, measured at 3.306814 bpw
• public highest-quality release: Laguna-S-2.1-Alis-MLX-Highest-Quality-4.892631bpw, measured at 4.892631 bpw
• the full checkpoint was exercised through 32K context; 1M context is a calculated support limit only and was not executed

The PR contains code and weight-free tests only. Full-checkpoint artifacts and quality evidence are published separately in the downstream release repositories above.

[JuanColilla]

Same PR like mine: #1602

Although I see yours was first, which one shall we close?