fix(deltanet): restore L2 norm (removal causes output collapse) + decay fix kept

Karpathy loop results on Qwen3.5-4B short prompts:
  Loop 1: decay formula aligned with llama.cpp → still fails (0/5)
  Loop 2: L2 norm removed → WORSE (doc QA also breaks)
  → L2 norm restored (REQUIRED for Qwen3.5)

The decay formula fix (sk before decay) is kept as it's
mathematically correct per llama.cpp reference.

Remaining suspect: Q scaling timing or state shape/layout mismatch.
The reference analysis found llama.cpp uses [S_v, S_v] square state
while quant.cpp uses [dk, dv] rectangular — this needs investigation.

Refs #95

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

Author: unamedkr

quant.h

@@ -13890,6 +13890,8 @@ static void deltanet_forward(tq_model_t* model, tq_state_t* s, int l) {
     float* K_all = s->delta_qkv + dn_kv * dk;
     float* V_all = s->delta_qkv + 2 * dn_kv * dk;
 
+    /* L2 normalization of Q/K: REQUIRED for Qwen3.5-4B.
+     * Removing this causes complete output collapse. */
     for (int h = 0; h < dn_kv; h++) {
         l2_normalize(Q_all + h * dk, dk);
         l2_normalize(K_all + h * dk, dk);