Port Phi-3 fused QKV/FFN from quant.h to libturboquant (enables Metal GPU)

[unamedkr]

Summary

Complete the Phi-3 architecture port from quant.h to src/engine/*.c so that quant-server (libturboquant) can serve Phi-3.5 with Metal GPU acceleration.

Current State

• quant.h: Phi-3.5 works perfectly (6.5 tok/s CPU NEON)
• libturboquant: Phi-3.5 crashes or produces garbage
• Workaround: quant-server-unified (compiles quant.h directly, no GPU)

What Needs Porting

Feature	quant.h	src/engine/	Status
Fused attn_qkv detection	✅	Partial	loader works, forward crashes
Fused ffn_up_gate detection	✅	Partial	loader works
LongRoPE NeoX rotation	✅	Written	untested
K/V projection skip for fused QKV	✅	Missing	root cause of crash
State allocation (xb2, hb sizing)	✅	Written	needs verification
BOS token handling	✅	Written	untested

Previous Attempt

PR #71 attempted this but caused SmolLM2 regression — the K/V projection skip interacted badly with existing code paths. The port needs to be done one component at a time with regression tests after each step.

Recommended Approach

1. Port K/V skip only → test SmolLM2 + Phi-3.5
2. Port FFN fused path → test
3. Port LongRoPE → test
4. Metal dispatch → benchmark

Impact

Metal GPU acceleration for Phi-3.5 could improve speed from 6.5 to ~15-20 tok/s on Apple Silicon (3-4B models benefit from Metal, unlike 1B).

Priority: P3

This is blocked by #85 (Single Source of Truth) — once that's done, porting is automatic.