model: add support for Mellum2 architecture

[joelfarthing]

Summary

Adds JetBrains Mellum2 support, including:

• mellum architecture registration and model type reporting
• Mellum tensor mapping/loading
• graph support for Mellum's mixed sliding-window/full-attention layer pattern
• per-layer SWA/full RoPE handling
• routed MoE expert tensor loading
• Mellum2 tokenizer pre-tokenizer support
• converter support for official JetBrains/Mellum2-12B-A2.5B-* weights
• converter metadata for Mellum sliding-window pattern and full-attention/SWA RoPE settings

This PR does not include MTP/speculative decoding, performance tuning, custom quantization work, or chat/server UX changes.

Validation

Official JetBrains Instruct source revision:

4ee5751ef73ac6ae5a65b76b092ffc7c3b9c60e3

Public branch for reviewer inspection:

https://github.com/joelfarthing/ik_llama.cpp/tree/mellum2-support

Reviewer/community convenience GGUFs generated from official JetBrains Instruct BF16 weights:

https://huggingface.co/ji-farthing/Mellum2-12B-A2.5B-Instruct-ik-llama-GGUF

These GGUFs were converted with this PR's ik_llama.cpp converter. They embed the stock JetBrains chat_template.jinja and include the Mellum RoPE/YARN metadata emitted by this PR. They are expected to run on current llama.cpp builds with Mellum2 support as well as this ik_llama.cpp branch.

Validation highlights:

• current origin/main rejects Mellum2 GGUFs with unknown model architecture: 'mellum'
• converted BF16 and Q8_0 include Mellum rope.freq_base, rope.freq_base_swa, YaRN scaling metadata, sliding-window size, and sliding-window layer pattern
• converted BF16 and Q8_0 embed the stock JetBrains Instruct chat template byte-for-byte
• Q8_0 passes ik_llama.cpp CUDA server long-code smoke with --cpu-moe
• Q8_0 passes current llama.cpp upstream CUDA server long-code smoke with --cpu-moe
• BF16 passes current llama.cpp upstream CUDA server long-code smoke with --cpu-moe
• long-code smoke responses stop normally, compile with python3 -m py_compile, and pass a functional topological-sort test including cycle detection
• llama-server and llama-quantize CUDA build targets are up to date
• converter/GGUF py_compile passes
• git diff --check passes

Notes / caveats

• The linked GGUFs are convenience artifacts for review and community testing; they are not performance or quality claims.

• The model's normal logits path uses the standard output head. MTP/draft-head support is not required for normal Mellum2 generation and is intentionally out of scope here.

• The uploaded BF16 is a reference conversion; Q8_0 is the practical reviewer/community download.

• No custom chat template is used. The embedded template matches JetBrains/Mellum2-12B-A2.5B-Instruct/chat_template.jinja.

• No performance claims are made.

• I have read the contributing guidelines

• Self-reported review complexity:

  • Low
  • Medium
  • High