[bri-prism]

What

• Extends the 8-rows-per-block one-column mat-vec geometry, which only PQ2_0 used, to Q8_0, Q4_0, Q2_0, Q4_K and Q6_K on Hopper when the matrix has at least 4096 rows (a multiple of 8). Default on; GGML_CUDA_MMVQ_TALL_EXT=0 restores the old PQ2_0-only behavior.
• Adds Gemma 4 to the prism.hadamard architecture whitelist.

Why

• Output heads and other very tall matrices in these types ran as one 128-thread block per row. A 262144 x 2816 Q8_0 head reached only about 1.6 TB/s on H200.
• Gemma 4 routes every folded weight (attention q/k/v/o, shared FFN, routed experts) through build_lora_mm / build_lora_mm_id, so the activation transform is applied correctly. Until now the loader refused folded Gemma 4 GGUFs.

How

• Same kernel, same K split and per-row reduction order, so every output row is bit-identical to the default geometry. Only the rows-per-block launch parameter changes.

Testing (H200)

• test-backend-ops MUL_MAT 1669/1669 and MUL_MAT_ID 1035/1035 pass on this branch (CUDA vs CPU).
• Greedy output is byte-identical with the switch on and off on several models.
• No change on PQ2_0 models, which already used the tall path (27B PQ2_0 decode 172.0 vs 172.1 tok/s).
• Decode: +3.8% on a 2B Q8_0 model and about +7% on a Gemma 4 MoE model with a Q8_0 head.
• Hopper only; other architectures are unchanged.