Add data sampling to the training pipeline

[jackapbutler]

For training on a mixture of natural language and chemistry datasets we want to be able to load more than 1 Hugging Face tokenised dataset into the training pipeline specifying some sampling metrics which define the composition of the final training dataset.

i.e.

data:
  paths: /fsx/proj-chemnlp/data/EleutherAI/pythia-1b/marianna13/chemrxiv .... <another>
  ratios: 25:75 # final dataset has a composition of 25% dataset A and 50% dataset B

See this issue for further conversation of possible implementations.