Clean up the logs, bump datasets and transformers

[dakinggg]

The noisiest logs were coming from an old datasets version. Bumping the datasets version ended up requiring bumping the transformers version due to a dependency quadrangle. So this PR is including the version bumps too (and associated workarounds).

• Run regression tests against this PR

llama2 finetuning, before and after, with and without flash attention enabled

• Verify that model saving and deployment handles safetensors (turned on by default in latest transformers)
• Run a before and after for the logs (before: clean-logs-baseline-EJoND1, after: clean-logs-local-after-slljhI)

Related transformers issues:

• Transformers 4.36 use_cache issue huggingface/transformers#28056
• ValueError: Flash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes. You passed torch.float32, this might lead to unexpected behaviour. huggingface/transformers#28052

[irenedea]

Can you also update streaming to 0.7.2? Actually we can do a separate PR for streaming and composer updates, wdyt? Though, I was thinking it might be nice to do it all at once for regression tests.

[irenedea]

looks good! had a suggestion to use a context manager for filtering cases that require clean up