add support for defined train split

[winglian]

fixes #653

a lot of people are coming in using Jeremy's tutorial and using the sql dataset which doesn't have a train split. This adds a way to specify a split to train on and adds a quick type that matches his example in https://github.com/fastai/minai/blob/main/llm_example.ipynb

dataset:
    - path: knowrohit07/know_sql
      type: context_qa.load_v2
      train_on_split: validation