fix: Qwen3.5-MoE uses standard RMS norm (not residual), fix garbled output

The Qwen3.5-MoE model stores full RMS norm weights (centered ~1.0), not
residuals. With residual_rms_norm=true, cake was adding 1.0 to the weights,
doubling the norm output and causing activations to explode through 40 layers.

Also removes debug logging from GPTQ dequant paths.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: evilsocket

Cargo.lock

@@ -129,7 +129,7 @@ impl Qwen3_5MoeConfig {
             head_dim: Some(tc.head_dim),
             partial_rotary_factor,
             linear_attn,
-            residual_rms_norm: true,
+            residual_rms_norm: false,
             use_qk_norm: false,
             pre_reshape_qk_norm: false,
             sliding_window: None,
@@ -212,7 +212,7 @@ mod tests {
         assert_eq!(c.shared_expert_intermediate_size, Some(8192));
         assert!(c.attn_output_gate);
         assert!(c.norm_topk_prob);
-        assert!(c.residual_rms_norm);
+        assert!(!c.residual_rms_norm);
         assert_eq!(c.model_prefix, "model.language_model");
         assert_eq!(c.head_dim, Some(256));
         assert_eq!(c.rope_theta, 10_000_000.0);