realizar: batched GEMV row-equal to the single-row kernels for Qwen3.5 (#5279 P3) #5303

Phase P3 of #5279. Builds on #5289 (Qwen35CudaModel::forward_batch_greedy) and #5293 (the serve batch engine).

At m > 1, forward_batch_greedy still runs one full forward per slot, so every projection streams its weights m times per step. On gx10 (Qwen3.5-4B Q4_K_M) four concurrent requests each decode at 24.9% of a lone request's rate. The V3 shape rule needs at least 50%.

Change.

• Measure, on the real 0.8B file, whether the existing batched GEMV kernels (batched_hw_dp4a_q4k_gemv_into, batched_q4k_gemv_into, batched_q6k_gemv_into) give each row bitwise the same output as gemv_dispatch on that row alone. This is the single-row path a lone request takes.
• Use a batched kernel for a qtype only where each row is bitwise equal. Otherwise add a batched kernel that keeps the single-row kernel's per-row accumulation order: it loads each weight block once and dots it against m rows.
• The Q8_1 activation buffer grows to hold m rows, once, when the model is built.

Done when.

• A test on the real file holds each batched GEMV used by the Qwen3.5 forward bitwise equal to gemv_dispatch row by row, for m in {1, 2, 4, 8, 16} and for every qtype the file uses (Q4_K, Q5_K, Q6_K, Q8_0).
• A qtype that cannot reach bitwise equality is named, with its measured max difference, and stays on the per-row loop.

Not in this phase. Wiring the kernels into the forward (P4).