Fix get_wikitext2 tokenization bug causing sequence length warning

[Mr-Neutr0n]

Summary

Fixes #2020

get_wikitext2 was concatenating up to 1000 dataset entries into a single string and tokenizing it all at once. This produced a 73K+ token sequence that exceeds the model's maximum sequence length, triggering the warning:

Token indices sequence length is longer than the specified maximum sequence length for this model (73218 > 2048). Running this sequence through the model will result in indexing errors

This fix changes the function to tokenize individual samples instead, consistent with how get_c4 and get_c4_new already work. Each sample is tokenized separately and retried with a different random sample if it is shorter than the requested seqlen.

Changes

• Removed concatenation of 1000 text entries into a single string
• Removed bulk tokenization of the entire concatenated text
• Added per-sample tokenization with retry loop (matching get_c4 / get_c4_new pattern)

Test plan

• Verify get_wikitext2(tokenizer, nsamples=128, seqlen=32, split="train") no longer produces the sequence length warning
• Verify the returned dataset has the correct number of samples and shape
• Verify quantization workflows using wikitext2 dataset still work correctly

[Mr-Neutr0n]

Friendly bump! Let me know if there's anything I should update or improve to help move this forward.