Add Meituan LongCat-2.0 support / Unsloth variants

## Summary Please consider adding support for Meituan LongCat-2.0 to Unsloth, either as Unsloth-hosted optimized variants once the weights are available and/or as explicit compatibility support in Unsloth tooling. Official announcement: https://longcat.chat/blog/longcat-2.0/ GitHub repository: https://github.com/meituan-longcat/LongCat-2.0 Hugging Face repository: https://huggingface.co/meituan-longcat/LongCat-2.0 API / platform docs: https://longcat.chat/platform/docs/ ## Why this would be useful LongCat-2.0 is a newly announced large-scale MoE model focused on agentic and coding workloads. The official repository describes it as a 1.6T total-parameter model with about 48B activated parameters per token, 1M-context training, and strong reported results on coding-agent benchmarks such as Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, and LongCatClawBench. Since Unsloth is one of the easiest ways for many users to run, quantize, fine-tune, or experiment with open models locally, adding LongCat-2.0 support would make this model family much more accessible once public checkpoint files are available. ## Current availability The official LongCat-2.0 GitHub repository and Hugging Face model repository are live. The repository currently notes that model weights are "coming soon", and the Hugging Face repository currently appears to contain the model card/assets rather than checkpoint files. I also checked the current Unsloth Hugging Face model listings through the Hugging Face API for Unsloth-owned models matching `LongCat` or `meituan`, and did not find current Unsloth-hosted LongCat variants. There is an older issue for `Longcat-flash` compressed-model support (#3608), but this request is specifically for the newly announced `meituan-longcat/LongCat-2.0` model. ## Suggested scope Useful additions could include one or more of: - Unsloth-hosted quantized variants for `meituan-longcat/LongCat-2.0` once checkpoint files are public. - Documentation or examples for loading, quantizing, or fine-tuning LongCat-2.0 through Unsloth if the architecture is compatible. - Confirmation of whether LongCat-2.0's MoE architecture, long-context setup, sparse attention, and related components are already supported, partially supported, or need model-specific work. - If full training support is not immediately feasible, an initial inference/quantization compatibility note would still be helpful. ## Practical starting point The Hugging Face repository is the canonical handle to track: `meituan-longcat/LongCat-2.0` Because the weights appear to still be pending, this could start as a compatibility tracking issue and become actionable once the model files/config are published. Thanks for considering it.