feat(ruvector-cnn): CNN contrastive learning + SIMD optimization fixes

[ruvnet]

Summary

This PR adds CNN contrastive learning capabilities to ruvector and fixes critical SIMD/backbone issues discovered during code review.

New Features

CNN Contrastive Learning Crate

• InfoNCE Loss (SimCLR-style) with temperature scaling and numerical stability
• Triplet Loss with configurable margin and hard/soft modes
• Augmentation Pipeline (requires augmentation feature): random crop, flip, color jitter, grayscale, blur
• MobileNet-V3 Backbone (Small/Large variants)
• SIMD Optimizations: AVX2, NEON, WASM SIMD128
• WASM Bindings (ruvector-cnn-wasm)
• npm Package (ruvector-cnn) with TypeScript definitions

Bug Fixes (Closes #251)

SIMD Functions Were Stubs

Implemented real SIMD for:

• conv_3x3_avx2, conv_3x3_avx2_fma - now processes 8 output channels simultaneously
• depthwise_conv_3x3_avx2 - now processes 8 channels at once
• global_avg_pool_avx2 - vectorized accumulation
• max_pool_2x2_avx2 - uses _mm256_max_ps
• conv_3x3_wasm, depthwise_conv_3x3_wasm - 4 channels with SIMD128

MobileNetV3 Forward Pass

• Fixed block processing (was completely skipped)
• Deprecated MobileNetV3Small/MobileNetV3Large in favor of unified MobileNetV3
• Fixed hardcoded channel counts

Dead Code

• Silenced unused field warnings with proper annotations
• Fixed undefined feature reference

Test Plan

• All 36 unit tests pass
• Doc-tests pass
• WASM crate compiles
• npm package ready

Files Changed

crates/ruvector-cnn/           # Core crate (36 tests)
crates/ruvector-cnn-wasm/      # WASM bindings
npm/packages/ruvector-cnn/     # npm package
docs/research/cnn/             # Research documentation
docs/adr/ADR-088-...           # Architecture decision

🤖 Generated with claude-flow

[ruvnet]

Additional Changes (Latest Commit)

Added more optimizations and documentation improvements:

New Features

• Winograd F(2,3) transforms — 2.25x faster 3x3 convolutions (36→16 multiplications per 2x2 output)
• π-calibrated INT8 quantization — Uses π-derived anti-resonance offsets to prevent bucket collapse
• 4x loop unrolling — AVX2 convolutions with 4 independent accumulators for better ILP

Documentation Updates

File	Changes
crates/ruvector-cnn/README.md	Practical intro, Quick Start examples, capabilities/performance tables, use cases
npm/packages/ruvector-cnn/README.md	Simpler language, step-by-step examples, troubleshooting
README.md	Added CNN image embeddings to root capabilities list

New Files

• simd/winograd.rs — Winograd transform matrices, filter caching, conv function
• simd/quantize.rs — π-calibrated INT8 quantization with AVX2 SIMD

All 36 unit tests + 4 doc tests pass.