Bug fixes

[danielhanchen]

1. Fixes Gemma 3N failing on torch 2.7 with out of bounds errors - [Bug] Gemma 3n inference fails on Windows #2941 Possibly [Bug] Finetune Gemma-3n throws canUse32BitIndexMath error #2940
2. Fixes GRPO state2 missing / NaN for vllm >= 0.9.1
3. Fixes mamba style training for Liquid, Falcon H1 etc
4. Fixes many other small bugs
5. Fixes In GRPO special tokens are removed if max_prompt_length is set, else not #2980 - special tokens and pad tokens not handled properly in GRPO
6. Fixes [Bug] Compile cache miss typing.Any #2874
7. Inference somewhat faster

Update Unsloth via:

pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth unsloth_zoo