PERF: revert prefetch attempt — no measurable benefit, restore clean gather

Round 8 (prefetch in gather loop) gave no measurable improvement on
Apple M1 Pro — within noise across 3 runs. The L1 prefetcher already
handles the strided pattern, and explicit __builtin_prefetch hints add
no value here. Reverting to keep the code clean.

Round 9 (strided per-position attention without gather) was considered
but rejected: would require either repeated query rotation per position
(slower) or a new traits ABI for pre-rotated single-block dot product
(invasive). Current gather + bulk attention is the local optimum for
this attention path on Apple Silicon.

Final honest numbers (3 runs, Llama 3.2 3B PPL eval):

  fp32         14.63 t/s   baseline
  turbo_kv_4b  13.57 t/s   −7.2%   ⭐ default
  turbo_kv_3b  13.13 t/s   −10.2%
  turbo_kv_5b  12.90 t/s   −11.8%  🏆 quality

35/35 tests pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

src/engine/tq_transformer.c

@@ -1650,11 +1650,16 @@ static void self_attn_forward(tq_model_t* model, tq_state_t* s, int l, int pos)
             for (int t = 0; t < attn_start; t++) atth[t] = -1e30f;
 
             /* Fast path: no post-norm, no high-res window, attention type
-             * supports the fused kernel (which is true for all turbo_kv_*). */
+             * supports the fused kernel (which is true for all turbo_kv_*).
+             *
+             * Round 9: still gather + bulk attention. Skipping gather with
+             * strided per-position attention turned out NOT to help —
+             * Apple Silicon's prefetcher handles the strided pattern fine,
+             * and the gather lets the CPU's L1 prefetcher walk through a
+             * contiguous block, which is cache-efficient.
+             */
             if (!needs_post_norm && !k_hr_active && traits->attention != NULL
                 && attn_start == 0) {
-                /* Gather quantized blocks for this kv_head into a contiguous
-                 * buffer (the layer-stride layout has interleaved kv heads). */
                 size_t head_block_bytes = s->quant_head_stride;
                 size_t pos_stride_bytes = (size_t)cache_n_kv_heads * head_block_bytes;
                 uint8_t* layer_base = (uint8_t*)s->quant_key_cache
@@ -1667,7 +1672,7 @@ static void self_attn_forward(tq_model_t* model, tq_state_t* s, int l, int pos)
                     memcpy(gather_dst + (size_t)t * head_block_bytes, src, head_block_bytes);
                 }
 
-                /* Single bulk attention call computes all seq_len scores */
+                /* Single bulk attention call (query pre-rotated inside) */
                 traits->attention(qh, s->quant_key_buf, atth, seq_len, head_dim);
 
                 /* Apply scale */