fix(perplexity): use all_special_tokens for GPT-2 tokenizer compatibility

[akaashsa]

GPT-2's slow tokenizer (GPT2Tokenizer) does not expose special_tokens_map_extended, causing an AttributeError when computing perplexity with batch_size > 1.

Additionally, the pad_token guard was conditioned on batch_size > 1, but the tokenizer is always called with padding=True, meaning pad_token is required even for batch_size=1 when multiple predictions are passed.

Fix:

• Replace special_tokens_map_extended.values() with all_special_tokens, which is defined on PreTrainedTokenizerBase and returns a flat list of strings for both slow and fast tokenizers.
• Remove the batch_size > 1 condition so pad_token is set whenever it is missing, matching the actual padding requirement.

Fixes #766