fix(tokenizer): extend decoding fix to cover Mistral no-spaces bug

[eeshsaxena]

Fixes #1822

When chatting with Mistral models, the output was missing spaces between words. This is because Mistral uses the same SentencePiece-backed tokenizer as LLaMA (both have the prefix-space issue documented in huggingface/transformers#31643).

The existing apply_decoding_fix only checked for 'LlamaTokenizer' in the tokenizer_class, but Mistral checkpoints report 'MistralTokenizer' or 'MistralTokenizerFast', so the fix was never applied for them.

Changes:

• Extended the condition to also check for 'MistralTokenizer' in the tokenizer_class string
• Used .get() instead of direct key access to avoid KeyError if tokenizer_class is missing
• Updated the comment to explain the root cause (SentencePiece prefix-space encoding)

File changed:

• litgpt/tokenizer.py

[eeshsaxena]

Closing to test the fix locally first and add a test case before resubmitting.