[inureyes]

Summary

Migrate OLMoE from its local SwitchGLU/SwitchLinear pair to the shared crate::models::switch_layers::SwitchGLU, then wire the one-branch fused single-token decode dispatch. OLMoE's expert intermediate (1024) is below MLXCEL_FUSED_MOE_MAX_DFF (4096), so the kernel dispatches on every single-token decode step and delivers a real throughput gain.

What changed

• src/models/olmoe.rs: removed local SwitchLinear enum (68 lines) and local SwitchGLU struct with forward/gather_sort/scatter_unsort impls (117 lines). No other module imported either type.
• SparseMoeBlock.experts retyped from the deleted local SwitchGLU to crate::models::switch_layers::SwitchGLU.
• SparseMoeBlock::from_weights: construction now calls crate::models::switch_layers::SwitchGLU::from_weights(weights, "{layer}.switch_mlp", group_size, bits). Standard gate_proj/up_proj/down_proj leaf names match the shared loader directly; the existing sanitize_weights helper already stacks per-expert weights under that prefix.
• Removed local impl SwitchGLU { from_weights } and impl SwitchLinear { from_weights } (40 lines). get_weight_copy and sanitize_weights are kept (still used by Attention and OlmoeModel::from_weights).
• SparseMoeBlock::forward: bare experts.forward + moe_weighted_sum dispatch replaced with the standard fused-kernel branch (mirror of qwen3_moe/phimoe).
• Added a narrow compile-time unit test confirming fused_moe_enabled() is callable from olmoe's scope.
• docs/benchmark_results/fused-moe-decode-kernel-design.md: wired-path count bumped from nine to ten; olmoe entry moved from the not-wired list to the wired list; not-wired list now contains only minimax.

Test plan

• cargo check --lib --tests --features metal,accelerate passes clean
• cargo clippy --lib --tests --features metal,accelerate -- -D warnings passes clean
• Real-checkpoint validation (mlx-community/OLMoE-1B-7B-Instruct-4bit) pending the orchestrator

Closes #302

[inureyes]

Rebased onto main (picks up the olmoe q_norm crash fix #316). Orchestrator validation on mlx-community/OLMoE-1B-7B-0125-Instruct-4bit (M1 Ultra): the fused dispatch is wired correctly and fires (fused-on differs from MLXCEL_FUSED_MOE=0 only within the documented f16 jitter class), and decode is perf-neutral (~250 tok/s both on and off; no gain at olmoe's top_k=8 / Dff 1024, and no regression). The migration removes the local SwitchGLU/SwitchLinear duplication. Note: olmoe still emits degenerate output on the gather_qmm baseline too, from a separate pre-existing defect (not q_norm, not the MoE path) tracked in #318; this PR neither causes nor worsens it. The fused wiring is parity-preserving relative to the baseline.