perf(core): add fused Qwen 3.5 vector kernels

[bsbodden]

Summary

• add SIMD SwiGLU and gate/softplus primitives
• add fused causal depthwise-convolution + SiLU with tap-major state
• cover scalar/Panama parity, aliasing, offsets, tails, and numerical stability

Verification

• ./gradlew --no-daemon spotlessCheck test
• Qwen3.5-0.8B controlled benchmark: 27/27 correct, 80.61 tok/s median decode, 1.649 s p95 end-to-end