Fix incorrect kv_head_idx computation breaks GQA/MHA head-to-KV-head mapping

[oyeong011]

Summary

kv_head_idx = head_idx / _nkvh is mathematically wrong. num_heads (the function parameter at both call sites of Attention::initialize_instructions) already equals heads_per_kv (= num_attention_heads / num_kv_heads), and head_idx is always a multiple of heads_per_kv (the starting query-head index of that kv-head's group). The correct divisor is num_heads, not _nkvh.

Symptoms by head ratio (A = num_attention_heads, K_h = num_kv_heads)

Ratio	Example	Symptom
MHA (A == K_h)	opt-125m (12/12), Llama-2-7B (32/32)	Every head silently reads/writes kv_head_idx=0 instead of its own slice — all heads collapse onto one KV head. No crash.
GQA, A/K_h < K_h	Llama-3-8B (32/8, ratio 4)	Some kv heads (e.g. 4 of 8) are never accessed. Silent, no crash — indices stay in-bounds by coincidence.
GQA, A/K_h > K_h	Qwen2.5-7B (28/4, ratio 7)	kv_head_idx exceeds num_kv_heads, out-of-range index crash (make_address: index is out of bound).

Verification

After the fix, kv_head_idx was logged across four models and matched the expected full range [0, num_kv_heads) in every case:

• opt-125m (12/12): 0..11
• Llama-2-7B (32/32): 0..31
• Llama-3-8B (32/8): 0..7 (previously only 0..3 were ever reached)
• Qwen2.5-7B (28/4): 0..3 (previously crashed)

Fixes #33.

Scope

This PR contains only the two-line fix. It does not include any of the research instrumentation (dequant engine, KV-separated metrics, quantization metadata modeling) built on top of it for the study that surfaced this bug.