get_dataset() doesn't require context_length, but the TatmMemmapDataset constructor does

[toizzy]

get_dataset calls TatmMemmapDataset() with **kwargs, and errors when optional argument context_length is missing

>>> from tatm.data import get_dataset, torch_collate_fn
>>> get_dataset("/dolma_path/dolma/tokenized/baseten-meta-llama-3-tokenizer/c4")                                                                                                                  
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/conda_env/lib/python3.12/site-packages/tatm/data/datasets.py", line 49, in get_dataset                                                                                                 
    return TatmMemmapDataset(
           ^^^^^^^^^^^^^^^^^^
TypeError: TatmMemmapDataset.__init__() missing 1 required positional argument: 'context_length'

[mbsabath]

This is somewhat expected behavior, as get_dataset is intended as a general interface to training ready dataset objects (of which there's currently only one, but there will likely be more). That said, we can think about better ways of checking for inputs, or figure out a plan for making sure that expected inputs are well documented, I'll leave this open until that gets cleared up