[feat][gpu] Q4 quantization, Metal GPU shaders, ANE kernel fusion, memory safety

Inference engine:
- Q4_0 quantization format (block_size=32, 20 bytes/block) with convert_weights.py
- Q4 AMX dequant-at-load path: same ~91 t/s as F32 (bandwidth-bound)
- Metal GPU compute shaders (matmul.metal): 15+ kernels including SIMD Q4
  matmul, fused gate+up+SiLU FFN, batched attention, batched prefill
- GPU path reverted from generate() due to dispatch overhead making it
  slower than CPU AMX for dim=896 (kept as dead code for larger models)
- ANE kernel fusion (GQA-aware QKV + Gate/Up) reduces 184->112 kernels,
  fits under 119 compile limit. Compiles but ane_eval fails on M4.

Memory safety:
- safe_malloc/qwen_calloc wrappers with OOM crash protection
- Moved per-layer malloc/free out of 24-iteration loop in prefill
  (was allocating/freeing up to 17MB buffers 168 times per request)
- Removed GPU weight upload and --gpu/--bench-gpu flags

Benchmark:
- Multi-format comparison: F16, Q8, Q4_AMX, Q4_Metal, 3 LM Studio models
- Wall-clock round-trip timing for fair cross-engine comparison

Made-with: Cursor

Author: dev-erik

.gitignore

@@ -25,6 +25,9 @@ training/test_*
 # Inference binaries and runtime data
 inference/qwen_ane
 inference/qwen05b.bin
+inference/qwen05b_f32.bin
+inference/qwen05b_f16.bin
+inference/qwen05b_q8.bin
 inference/.venv/
 inference/benchmark_results.json
 
@@ -59,6 +62,7 @@ web/
 training/tinystories_data00.bin
 training/ane_stories110M_ckpt.bin
 *.bin
+*.metallib
 !training/download_data.sh
 
 # Secrets / env