readme: clarify LFW methodology + soften encryption claims (refs face…

…x-engine#6, facex-engine#7)

Recognition accuracy table now explicitly states:
- mean across 10 folds (not best fold)
- standard deviation per variant
- 5-point ArcFace alignment is required; un-aligned input drops ~25 points
- link to training/scripts/lfw_eval.py for reproduction

Encryption tagline rewritten to be honest: weights are AES-256-GCM
ciphertext on the wire, but this is *not* DRM — the decrypted bytes
sit in the WASM heap and are extractable. The actual value is
casual-scrape friction, per-customer key revocation for SaaS, and
audit trail at the key-issuing endpoint. Pointer to the wiki page
which has the threat model + Express / FastAPI integration recipes.

Add golden test, download script, optimize WASM to 7ms

- Rename go/facex/fastface.go → facex.go, update package name
- Add download_weights.sh for fetching model from GitHub Release
- Add tests/golden_test.c — deterministic forward pass verification
- Add `make test` target
- GEMM K-unroll=2 in packed matmul (+3% WASM speed)
- Optimized WASM build: 14ms → 7ms median (AVX2 SIMD emulation)