Function load_moe_minimax 

pub fn load_moe_minimax(
    store: &WeightStore,
    layer_prefix: &str,
    num_experts: usize,
    gpu: &dyn GpuBackend,
    config: &ModelConfig,
    variant: Nvfp4Variant,
    absmax_k: KernelHandle,
    quantize_k: KernelHandle,
    stream: u64,
) -> Result<MoeWeights>

2026-09-25: Load a MiniMax M2 MoE block.

Differences from load_moe_no_shared:

• Prefix {layer}.block_sparse_moe instead of {layer}.mlp.
• Experts use w1/w2/w3: w1 = gate_proj, w2 = down_proj, w3 = up_proj.
• e_score_correction_bias goes into MoeWeights.correction_bias, widened to F32 when it is BF16.

Like load_moe_no_shared it builds a zero-filled shared expert (packed codes, scales and weight_scale_2 all 0), which contributes 0.