feat: add Motif-3 model architecture support

[ajaxdude]

Summary

Add support for the Motif-3 model architecture (314B.A13B MoE) from AesSedai/Motif-3-GGUF.

Architecture features

• GDLA (Grouped Differential Latent Attention): Attention variant with per-head lambda projections and latent KV cache (MLA-style)
• Grouped PolyNorm MoE: PolyNorm activation with grouped expert routing and shared expert fallback
• mHC (Manifold-constrained Hyper-Connections): Residual stream expansion with Sinkhorn-normalized gating tensors (h_pre, h_post, h_res)
• Interleaved SWA + YaRN: Sliding window attention with YaRN rope scaling

Changes

File	Change
llama-arch.{h,cpp}	MOTIF3 arch enum, 8 KV fields, 14 tensor fields with info
llama-model.{h,cpp}	26 layer tensors, model constructor mapping, rope type
llama-hparams.h	PolyNorm eps/out_scale/clamp, mHC mult/iters/post_coeff, GDLA MLA dims
llama-graph.cpp	MLA KV cache integration (K-only base cache), buffer null checks
llama-kv-cache-iswa.{h,cpp}	SWA subcache hparams override to disable MLA on SWA layers
llama-context.cpp	Graph node allocation ceiling for motif3
llama-quant.cpp	Motif-3 tensor quantization exclusions (PolyNorm, mHC, GDLA weights)
llama-vocab.cpp	<|endofturn|> end-of-turn token
gguf-py/gguf/constants.py	Enum, tensor names, allowed/skip tensor lists
conversion/__init__.py	MotifForCausalLM → motif3 mapping
conversion/motif3.py	PyTorch → GGUF converter (373 lines)
src/models/motif3.cpp	Full architecture implementation (907 lines)
src/models/models.h	llama_model_motif3 struct declaration

Based on

Upstream PR: ggml-org/llama.cpp#26298

Testing

• Verified against upstream PR model: Add support for Motif 3 Beta  ggml-org/llama.cpp#26298: all added lines present; new files conversion/motif3.py and src/models/motif3.cpp byte-identical to upstream
• cmake -B build -G Ninja + ninja llama: compiles clean (CPU build)
• cmake -B build-vk -G Ninja -DGGML_VULKAN=ON + ninja llama: compiles clean (Vulkan build; libggml-vulkan.so built)
• Converter unit-tested: all tensor mappings, wkv_b MLA split, mHC alpha grouping, and 33 arch-specific GGUF parameters verified with mocked tensors/writer
• Picked up automatically via file(GLOB LLAMA_MODELS_SOURCES "models/*.cpp") in src/CMakeLists.txt
• Full inference test not run: smallest quant (IQ2_S) is ~108 GB