perf(moe): wire phimoe to the fused decode-MoE kernel

[inureyes]

Summary

• Removes PhiMoE's local SwitchLinear / SwitchGLU (plus gather_sort / scatter_unsort helpers) and retypes SparseMoeBlock.experts to the shared crate::models::switch_layers::SwitchGLU.
• Wires the single-token fused decode dispatch in SparseMoeBlock::forward, matching the pattern used by qwen3_moe, gemma4, and lfm2.
• Keeps sanitize_weights intact so unstacked experts.{i}.w1/w2/w3 community checkpoints still stack correctly.

What changed

• src/models/phimoe.rs: removed ~190 lines of local SwitchLinear / SwitchGLU; retypes experts field; updates SparseMoeBlock::from_weights to call shared SwitchGLU::from_weights(weights, prefix+".switch_mlp", args.group_size(), args.bits()); adds fused dispatch block in SparseMoeBlock::forward.
• src/models/phimoe_tests.rs: new file with four config-parsing tests (4-bit checkpoint, head_dim, num_kv_heads, unquantized defaults).
• src/models/mod.rs: registers phimoe_tests.
• docs/benchmark_results/fused-moe-decode-kernel-design.md: moves phimoe to the wired models list; count seven -> eight.

Weight-loading parity note

Pre-stacked checkpoints (mlx-community/Phi-3.5-MoE-instruct-4bit) have weights under block_sparse_moe.switch_mlp.{gate,up,down}_proj.{weight,scales,biases}. The shared SwitchGLU::from_weights reads exactly those keys when called with prefix {layer}.block_sparse_moe.switch_mlp, so loading is identical to the old local implementation. The sanitize_weights stacking path for unstacked community checkpoints is untouched.

Test plan

• cargo check --lib --tests --features metal,accelerate passes
• cargo clippy --lib --tests --features metal,accelerate -- -D warnings passes (zero warnings)
• Real-checkpoint decode validation against mlx-community/Phi-3.5-MoE-instruct-4bit pending orchestrator (part of epic epic(moe): wire the remaining unwired MoE families to the fused decode-MoE kernel #307)

Closes #303

[inureyes]

Rebased onto main to pick up the Dff guard (#311). phi-3.5-moe's expert intermediate is 6400, above the default MLXCEL_FUSED_MOE_MAX_DFF (4096), so the kernel declines and decode stays on gather_qmm. M1 Ultra validation: guard-on 71.8-72.3 tok/s == MLXCEL_FUSED_MOE=0 baseline 72.6 tok/s (falls back, no regression); MLXCEL_FUSED_MOE_MAX_DFF=99999 re-arms the fused path at 68.8 tok/s (the -5.9% the guard avoids). The migration removes the local SwitchGLU/SwitchLinear duplication; the dispatch arms only if the bound is raised.