feat(models): add Mellum 2 hybrid-attention MoE text model

[inureyes]

Summary

Add JetBrains' Mellum 2 (model_type: mellum), a sliding/full hybrid-attention sparse-MoE code model, ported from mlx-lm (https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/models/mellum.py). The architecture is a composition of primitives mlxcel already has (QK-RMSNorm attention, softmax-routed MoE, mixed KVCache/RotatingKVCache, YaRN RoPE), wired into a new family loader.

What changed

• src/models/mellum.rs (new): config (serde), QK-RMSNorm Attention, softmax-routed SparseMoeBlock (top-k via argpartition, norm_topk_prob, routed through the shared SwitchGLU), an optional dense MLP path driven by mlp_layer_types, the pre-norm DecoderLayer, MellumModel, sanitize, make_caches, and the MellumWrapper that owns the mixed caches and implements LanguageModel.
• src/models/mellum_tests.rs (new): config parse, YaRN attention-factor derivation, sanitize expert-stacking, per-layer cache selection, and tied/untied head handling.
• src/models/detection.rs + src/models/detection_tests.rs: detect "mellum" -> ModelType::Mellum, with a detection test.
• src/models/mod.rs: mellum module + re-exports, the ModelType::Mellum variant, and its ALL_MODEL_TYPES / metadata() / arch-completeness entries, plus the mellum_tests module.
• src/loaded_model.rs: LoadedModel::Mellum variant + delegate_language_model! arm.
• src/model_metadata.rs: config-backed registration entry.
• src/distributed/tensor_parallel/inference.rs: fallback_architecture arm.
• docs/supported-models.md: Mellum entry.

Design notes

• Per-layer RoPE. rope_parameters is keyed by layer type. full_attention layers use YaRN (base 500000, factor 16) built once and shared per full layer; the derived mscale equals the config's explicit attention_factor (1.2772588722239782). sliding_attention layers use default RoPE. This mirrors mlx-lm initialize_rope / YarnRoPE.
• Hybrid caches. layer_types[i] selects a windowed mask + RotatingKVCache(max_size=sliding_window) for sliding layers and a causal mask + KVCache for full layers; the wrapper owns the mixed caches (supports_batching=false).
• QK-norm. q_norm/k_norm are RMSNorm(head_dim) applied to the reshaped per-head q/k before transpose and RoPE.
• MoE routing. Precise (f32-accumulation) softmax_precise over the router logits (matching upstream precise=True), top-k via argpartition, scores gathered with take_along_axis, normalized when norm_topk_prob, experts through the shared SwitchGLU (single-token fused kernel with the gather fallback).
• Untied head. tie_word_embeddings drives Option<lm_head>; sanitize pops a tied lm_head.weight and from_weights projects through embed_tokens.as_linear when tied. This checkpoint is untied.
• Expert stacking. sanitize stacks per-expert experts.{e}.{gate,up,down}_proj.* into switch_mlp.{proj}.* in place; the shared SwitchGLU also stacks lazily, so the weight-route load path works without an explicit pass.
• Attribution. mlx-lm model ports are already covered by the top-level NOTICE; the file carries the standard header plus an upstream URL reference, matching the existing mlx-lm ports (e.g. gpt_oss.rs).

Test plan

• cargo check --lib --tests --features metal,accelerate
• cargo clippy --lib --tests --features metal,accelerate -- -D warnings
• cargo fmt --check -p mlxcel
• cargo test --lib --features metal,accelerate -- models::mellum (6 pass)
• cargo test --lib --features metal,accelerate -- models::detection (detection + new Mellum case pass)
• arch-completeness tests (every_variant_is_registered_for_arch, all_model_types_covers_every_variant) pass

Real-model validation

Validated on the real JetBrains/Mellum2-12B-A2.5B-Base checkpoint (24GB bf16, 12B total / 2.5B active, 64 experts top-8) on M1 Ultra (Apple GPU / Metal): loaded in ~5.8s at 22.63 GB bf16 resident, greedy decode ~62-65 tok/s. mlxcel generate produced correct, coherent code on two prompts: a def quicksort(arr): completion that emitted the canonical pivot / left / middle / right recursive implementation, and a def binary_search(sorted_list, target): completion that emitted a correct low/high/mid loop. Degenerate output would have exposed a YaRN, QK-norm, or hybrid-attention bug, so this confirms the model graph. The router softmax was then upgraded to softmax_precise for exact parity with upstream and re-validated to still produce coherent output.

Closes #361