feat: dense arch pack for remaining Llama-family architectures (Mistral/Ministral, ExaOne, InternLM, Baichuan, GLM4, and more) #499

Part of #493

Context

The remaining Llama-family architectures are structurally close to Llama with minor deltas. Target set: Mistral variants / Ministral3, ExaOne / ExaOne4, InternLM2 / InternLM3, Baichuan, Apertus, SolarOpen, GLM4, Ernie4.5 (dense), Hunyuan (dense), Seed-OSS, MiMo, Step3.5.

Once the shared attention core (#494) lands, each is a config flag plus a small forward delta.

Scope

• Add each architecture via Config::from_json flags and the shared forward.
• Per-family weight_names ordering in iree.rs.
• Files: emitter/config.rs, emitter/model.rs, iree.rs.

Acceptance criteria

• Each listed architecture loads and generates via the CLI MLXCEL_BACKEND=xla path and serves via mlxcel-server.
• Single-sequence path is token-exact vs an HF fp32 oracle on a real checkpoint.
• Serve path is reference-exact via xla_batch_bench.

Dependencies

#494 (shared attention core). Uses #496 (validation harness) when available.