feat(models): port DeepSeek-V4 backbone (MLA + group-limited MoE)

[inureyes]

Summary

Adds DeepSeek-V4 support by porting the backbone only. DeepSeekV4Model reuses the DeepSeek-V3 stack (MLA attention, group-limited sigmoid-routed MoE, unclamped shared-expert DenseMLP, and V3's multi-token-prediction layer handling) as a thin wrapper: DeepSeekV4Config is parsed independently and mapped onto DeepSeekV3Config via to_dsv3_config, mirroring the existing glm_moe_dsa wrapper over DeepSeek V3.2. Config defaults track the mlx-vlm deepseek_v4 reference.

A deepseek_v4 checkpoint previously errored with "Unsupported model type"; it now detects, loads, and runs on the V3-equivalent backbone.

Scope

Backbone only, exactly as the issue directs. The two V4-specific features are deferred to separate follow-up issues and are intentionally not implemented here:

• HiSA (hierarchical sparse attention)
• the shared-expert SwiGLU swiglu_limit clamp

Their config fields (head_dim, swiglu_limit, num_nextn_predict_layers) are parsed for forward-compatibility but not applied by the backbone.

What changed

• src/models/deepseek_v4.rs (new): DeepSeekV4Config + DeepSeekV4Config::to_dsv3_config + DeepSeekV4Model (wraps DeepSeekV3Model).
• src/models/deepseek_v4_tests.rs (new): config default and V3-mapping unit tests.
• tests/deepseek_v4_parity.rs (new): reference-config parity + a synthetic forward-equivalence proof that the V4 backbone is bit-identical to DeepSeek-V3.
• Registration: ModelType::DeepSeekV4 and the deepseek_v4 detection.rs arm; LoadedModel::DeepSeekV4 + delegation; model_metadata.rs config-backed entry; tensor_parallel/inference.rs fallback arch string; ALL_MODEL_TYPES, metadata(), and the all_variants! registry list in mod.rs.
• docs/supported-models.md: DeepSeek-V4 backbone row.

Test plan

• cargo fmt
• cargo check --lib --tests
• cargo clippy --lib --tests -- -D warnings
• cargo test --lib deepseek_v4 (5 passed)
• cargo test --test deepseek_v4_parity (2 passed, incl. forward-equivalence vs DeepSeek-V3)
• cargo test --lib -- models::detection metadata_tests model_metadata (registry completeness, 19 passed)
• Real-checkpoint validation deferred (no public deepseek_v4 export yet)

Closes #523

[inureyes]

Closing without merging, per maintainer decision. This PR ported DeepSeek-V4 as a thin wrapper over DeepSeek-V3. Investigation against the in-tree reference (references/mlx-vlm/mlx_vlm/models/deepseek_v4/) and the real DeepSeek-V4-Flash checkpoint showed the real deepseek_v4 architecture is genuinely novel (HyperConnections, pooled-KV Compressor, HiSA Indexer, hash-routed MoE with sqrtsoftplus, MultiLinear output projection), not a V3 variant, so a V3 wrapper cannot load or run a real V4 checkpoint. Issue #523 is re-scoped to a faithful port of the real architecture. This PR's diff is preserved here for reference.