model: add Nanbeige4.2 (Looped Transformer) support

[Andgihat]

Adds support for the Nanbeige4.2 architecture (Nanbeige/Nanbeige4.2-3B, Apache-2.0) — a Looped Transformer: the physical decoder layers are reused num_loops times, so effective depth grows without adding parameters (22 physical layers × num_loops=2 → 44 effective).

Rebased onto current main, so the diff is a single clean commit.

What's included

• New arch LLM_ARCH_NANBEIGE + KV keys nanbeige.num_loops / nanbeige.skip_loop_final_norm.
• Graph builder src/models/nanbeige.cpp: standard RMSNorm + GQA + SwiGLU decoder with the loop unrolled over the logical (n_layer_all) layers. Physical weights are shared across loops via pointer copies (layers[i + j*n_phys] = layers[i]); each logical slot keeps its own KV index. An optional loop-boundary output_norm is applied between passes unless skip_loop_final_norm is set.
• gguf-py constants + writer keys and a conversion/nanbeige.py converter.
• Pure graph/arch code — no new ggml kernels.

load_arch_hparams expands the logical layer count via the existing n_layer (physical) / n_layer_all (logical) fields and replicates the per-layer arrays (n_head, n_head_kv, n_ff, is_swa_impl, is_recr_impl) across the loop copies, so the graph loop, KV cache and out-ids all size correctly with no changes to the base hparams API.

Verification

Built for Windows/CUDA 12.8 (sm_120) on top of current main and compared token-for-token against the authors' reference branch Nanbeige/llama.cpp@nanbeige42 on the same BF16 GGUF with greedy decoding — outputs are identical. Runs correctly with KVarN KV-cache (e.g. -ctk kvarn3 -ctv kvarn3) and long context.

GGUF quants (imatrix + KLD-guided mixed precision) are available at Andgihat/Nanbeige4.2-3B-GGUF.

[Anbeeld]

Does this upstream merge achieves the same thing? ggml-org@b77d646

[Andgihat]

I'll take a look now

[Andgihat]

Yes — b77d6467 (upstream ggml-org#25994) is the same implementation, merged by the model authors themselves. Same loop mechanism (logical-layer expansion + weight sharing layers[i + j*n_phys] = layers[i]), same per-layer array replication, same output_norm loop boundary gated by skip_loop_final_norm, and the exact same KV keys (nanbeige.num_loops / nanbeige.skip_loop_final_norm). No custom kernels on either side, so it also works with KVarN once master is merged in.

So this PR is redundant — merging upstream master is the cleaner path. Closing it. The GGUF quants at Andgihat/Nanbeige4.2-3B-GGUF are key-compatible and load on stock llama.cpp too.