Use Split-Scaling to avoid fp16 overflow:

Overflow Pattern:
` SiLU(w1) × w3 → [Inf!] → w2 → SLN`

Split Scaling Fix:
SiLU(w1) × (1/8) ──┐
├→ Mul_1 → w2 → SLN (no overfolw)
w3 × (1/16) ─────┘

Math: Mul_1 = SiLU(w1)/8 × w3/16 = SiLU(w1)×w3/128