fix(models/glm4-moe): fuse separate switch_mlp gate_proj/up_proj experts at load

[inureyes]

Follow-up hardening for #530: the merged GLM-4V MoE port fails to load its real checkpoint. mlxcel generate -m /path/to/glm-4.5v-4bit ... aborts with Failed to load GLM-4V MoE text model: Weight not found: model.layers.1.mlp.switch_mlp.gate_up_proj.weight.

Root cause

The GLM-4 MoE text loader expected a single fused switch_mlp.gate_up_proj expert stack. Real GLM-4(.5)V MoE checkpoints (converted by mlx-vlm) instead ship the routed experts as two already expert-stacked tensors, switch_mlp.gate_proj and switch_mlp.up_proj. Confirmed by reading the checkpoint's 12 *.safetensors headers directly (its model.safetensors.index.json is stale/pre-quant and was ignored):

language_model.model.layers.1.mlp.switch_mlp.gate_proj.weight  U32  [128, 1408, 512]
language_model.model.layers.1.mlp.switch_mlp.gate_proj.scales  BF16 [128, 1408, 64]
language_model.model.layers.1.mlp.switch_mlp.gate_proj.biases  BF16 [128, 1408, 64]
language_model.model.layers.1.mlp.switch_mlp.up_proj.{weight,scales,biases}   (same shapes)
language_model.model.layers.1.mlp.switch_mlp.down_proj.weight  U32  [128, 4096, 176]
language_model.model.layers.1.mlp.shared_experts.{gate,up,down}_proj.{weight,scales,biases}
language_model.model.layers.1.mlp.gate.weight                  BF16 [128, 4096]
language_model.model.layers.1.mlp.gate.e_score_correction_bias BF16 [128]

Fix

SwitchGLU::from_weights (in src/models/glm4_moe.rs) now prefers a pre-fused gate_up_proj when present and otherwise fuses the separate tensors at load via the new SwitchLinear::from_separate_gate_up.

• Axis / order: concatenate gate-then-up along the expert output-feature axis, i.e. axis 1 of [num_experts, out_features, in_features] ([128, 1408, ...] + [128, 1408, ...] -> [128, 2816, ...]). Gate is the first half, matching the [0:I) = gate, [I:2I) = up split already done in SwitchGLU::forward.
• Quant-aware: 4/8-bit weights pack along the input axis (last dim: 4096 -> 512 U32), so concatenating along the output axis never splits a packed word. The matching per-output-row scales/biases concatenate on the same axis. No dequantization is needed; the packed U32 words are carried through intact.
• down_proj, shared_experts, and the router gate (+ e_score_correction_bias) already match the checkpoint and load unchanged.

Analogous in spirit to the qwen3.5-moe per-expert -> fused fix in #519 / PR #587 (src/models/qwen3_5.rs), but in the opposite direction: GLM's SwitchGLU consumes a fused gate_up_proj, so this fuses rather than splits. The fusion order/axis were cross-checked against references/mlx-vlm/mlx_vlm/models/glm4v_moe/language.py, whose sanitize stacks gate_proj/up_proj/down_proj separately (the layout this checkpoint ships).

Because the fix lives in the shared Glm4Moe::from_weights path, both the standalone Glm4MoeModel and the Glm4vMoeTextModel VLM backbone accept the separate layout, and the pre-fused layout still loads (backward compatible).

What changed

• src/models/glm4_moe.rs
  • SwitchLinear::from_separate_gate_up: quant-aware fusion of separate gate_proj/up_proj (weight + scales + biases) along axis 1.
  • SwitchGLU::from_weights: use the pre-fused stack if present, else fuse the separate tensors.
  • get_weight_ref helper (borrow without copying, for in-place fusion).
  • Two unit tests: fuses_separate_switch_mlp_gate_up_experts (separate -> fused, asserts [E, 2I, ...] shapes for weight/scales/biases and unchanged down_proj) and loads_prefused_switch_mlp_gate_up_experts (backward compat).

Test plan

• cargo test --lib glm4 (9 passed, including the 2 new tests)
• cargo check --lib --tests
• cargo clippy --lib --tests -- -D warnings (clean)
• cargo fmt --check (clean)

Real-model GPU validation is intentionally deferred to the orchestrator (this is a 106B MoE; checkpoint was inspected read-only, not loaded here).