Upsampling: Statistical biasas of distribution of dataset

[michaelfeil]

I think there are some statistical biases in this implementation for long context engineering.

Concern 1: For upsample mode, some datasets groups get filtered when their capacity is maxed out. e.g for --down_sample_mode=upsample_code_arxiv_book, the code, arxiv and book datasets will be mostly at the end of our created syntetic dataset.

Concern 2: Start token_id 1. With the llama-tokenizer, when a single passage is tokenized, it is started by <s> or token_id1. When concetenating different pretokenized texts, its not the same result as if the strings are added and then tokenized together.