ruvector-cnn: SIMD optimizations were stubbed, backbone forward() was broken

[ruvnet]

Summary

Deep code review revealed critical issues in ruvector-cnn:

🚨 Critical Issues Found

1. SIMD Functions Were Stubs (Calling Scalar)

The following "SIMD-optimized" functions were actually calling scalar fallbacks:

Function	Platform	Status Before
conv_3x3_avx2	x86	❌ Scalar fallback
conv_3x3_avx2_fma	x86	❌ Scalar fallback
depthwise_conv_3x3_avx2	x86	❌ Scalar fallback
global_avg_pool_avx2	x86	❌ Scalar fallback
max_pool_2x2_avx2	x86	❌ Scalar fallback
conv_3x3_wasm	WASM	❌ Scalar fallback
depthwise_conv_3x3_wasm	WASM	❌ Scalar fallback

Impact: Users expecting SIMD acceleration were getting scalar performance.

2. MobileNetV3Small/Large Forward Pass Broken

// This was the actual code - blocks were not processed!
for _block in &self.blocks {
    // Each block would process here
    // For simplicity, we just pass through  <-- STUB!
}

Impact: Network produced incorrect outputs (stem only, skipping all blocks).

3. Dead Code Warnings

• Unused momentum field in BatchNorm
• Unused rng field in ContrastiveAugmentation
• Unused rand::Rng import
• Undefined parallel feature reference

✅ Fixed In This PR

1. Implemented real AVX2 SIMD for all 5 stubbed functions
2. Implemented real WASM SIMD for 2 stubbed functions
3. Deprecated MobileNetV3Small/Large (recommend unified MobileNetV3)
4. Implemented actual block processing in forward() methods
5. Fixed all dead code warnings

Test Results

All 36 tests pass ✅

---

Discovered via 4-agent parallel code review swarm