[ENHANCEMENT] Training with variable sequence length

[YuvalAtir]

I'd like to train llama3.1 model with variable sequence length as mentioned in Meta's paper. That is, to start the training with one sequence length and after a number of consumed tokens switch to a larger sequence length and continue training.

Several approaches in my opinion:

1. Train on dataset A with sequence length A for some iterations and save checkpoint. Continue training from checkpoint with a dataset B which has sequence length B. At runtime seq_length will be fixed and can be taken from args. This will require minimal changes.
2. Use BlendedMegatronDataset which can be used to build a dataset from several datasets. Currently, according to BlendedMegatronDatasetConfig all datasets share same seq_len. This could be changed to make seq_len a list instead of an int. The high level dataset will have variable sized samples. Some samples will come from dataset A with seq_len A and some will have seq_len B from dataset B. In this approach seq_length will not be fixed. It could be taken from a global function that returns the seq_length based on the index of the sample or the length of the sample.
3. A low level dataset could be comprised of variable sequence lengths. It will include a vector with the changes in sequence length. In runtime the sequence length would be taken from sample. At runtime sequence length is needed in several functions. In runtime seq_length will not be fixed, as in previous approach.

Any thoughts?