Llama3 pre-context parallel dataloader changes

[pstjohn]

Makes a number of updates in preparation for llama3 context-parallel training. It's still not currently working, need to further update the model to handle the cu_seq_lens_q_padded kwargs and would like to add a single-GPU CP test that uses BSHD inputs to at least exercise this code in CI.

This PR:

• Only materializes the dataloader on the cp_rank=0, and returns None on other ranks.
• Uses the scatter operation in the dataloader to synchronize StopIteration exceptions
• Adds tests for the CP dataloader on 1 and 2-gpu machines
• moves llama3 to use DLCM data as the sanity dataset, turns off some genome collation options by default. This is larger than the dummy sequences currently used in training, and will make sure we can fill a few batches in CP testing. We may want to revert this eventually once we're done bringing up llama3; since it does trigger the tokenizer download during testing.
• removes lazy tokenization from llama3, this wont work. See https://nvidia.slack.com/archives/C074Z808N05/p1767818883160949
• starts adding CP files for llama3

Closes BIO-11

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.