[bri-prism]

Depends on #413 (stacked on gemma/tall-ext; the diff here includes its two commits until it merges).

What

• Fuses the end of each Gemma 4 MoE layer into one kernel: the three RMS_NORM+MUL(+ADD) groups and the layer-scalar MUL.
• Optionally folds in the next layer's attn_norm (level 2) and the q8_1 quantization of that normalized activation for the next attention mat-vecs (level 3).
• GGML_CUDA_TAIL_FUSION=0/1/2/3 selects the level (default 0, off).
• GGML_CUDA_TAIL_VERIFY=1 also runs the unfused nodes and compares the outputs bit-for-bit. It is a debug mode and keeps a 128 MB scratch while enabled.

Why

• Decode on these models is one long chain of small dependent kernels. The four tail kernels are serial single-block norms on that chain, so removing them shortens every token.

How

• One block per row (ncols <= 3072), with the same per-thread partial sums, reductions and expression forms as the unfused kernels, so the result is bit-identical.
• Adds an ffn_res node name in gemma4.cpp so the pattern can be matched.
• Level 3 writes the q8_1 rows with the same arithmetic as quantize_q8_1. It is only used when every consumer is a plain MUL_MAT with the AOS layout.
• When the next attn_norm feeds a Hadamard-rotated input, levels 2/3 step aside so the existing norm+FWHT fusion still applies.

Testing (H200, this branch)

• Verify mode: 0 mismatches at levels 1 and 3, 64 fused groups each, on three Gemma 4 GGUFs.
• Greedy output is byte-identical off/on, repeated, on the same three and on a PQ2_0 Qwen3.5 27B.
• test-backend-ops RMS_NORM 51/51, NORM 50/50, ADD 99/99 and MUL 91/91 pass.
• Decode, with cuda: tall one-column mat-vec for Q8_0/Q4_0/Q2_0/Q4_K/Q6_K on Hopper, and allow folded Gemma 4 weights #413: about +16-19% over both switches off on Gemma 4 MoE GGUFs, of which about +7-8% is level 1 alone and up to about +12% is level 3 on unrotated models.