Many bug fixes

[danielhanchen]

Solves many issues. For Colab / Kaggle, please restart and run all.
For local machines, please do:

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

The main fixes are:

1. Fixes gibberish output for GRPO when not using vLLM
2. Mistral should work on GRPO
3. Torch 2.6 fixes
4. Various bug fixes and performance enhancements

Solves these issues:

1. NameError: name 'bias' is not defined [FIXED] #1895
2. modules_to_save doesn't work with PyTorch 2.6 #1882
3. The V100 ran GRPO training with incorrect output. #1888 Matrix mismatch runtime error #1875 GRPOTrainer example works with trl but generate "noise" with unsloth #1844 GRPO training error #1836 unsloth=2025.2.15 training result is werd when torch=2.6 #1825 GRPO without VLLM? #1810 - fixed all gibberish output for GRPO - apologies on the issue.
4. Slow Import Times (60+ seconds) When Using Unsloth via Entry Points #1859 Partially solved
5. UnslothTrainer applies ChatML template although passed train dataset is pre-tokenized and contains 'input_ids' field #1843
6. Llama AttributeError: 'bool' object has no attribute 'all_special_tokens' #1809
7. New GRPO doesnt support models besides LLAMA - (Mistral) #1790
8. No label_names provided for model class PeftModelForCausalLM. Since PeftModel hides base models input arguments, if label_names is not given, label_names can't be set automatically within Trainer. #1788