Fix usage of fract_base_dir and val_split

[michaelmckinsey1]

• Fixes setting fractal_dir. This was being ignored before
• Fixes setting val_split. This was not creating a new dataset because it wasn't in the hash
• Undoes val_batch_size. We should use the same for train and val

[PatrickRMiles]

Why do we want to be able to change the val_split from 75/25?

[michaelmckinsey1]

Why do we want to be able to change the val_split from 75/25?

I changed it to 30 because the default number of samples in the validation dataset with our current config is 60 at 25% split. At a 30% split it is 72. The target is 64, since that allows for a global batch size up to 64 instead of 32 (powers of 2 because of sharding).

We could also make it 27% to satisfy this constraint, but I just rounded up.

[PatrickRMiles]

Just one thing to discuss. Looks like github is having issues currently. Let me add that review comment back.

[PatrickRMiles]

In get_dataset.py, why include fract_base_dir in the dataset hash?

[michaelmckinsey1]

In get_dataset.py, why include fract_base_dir in the dataset hash?

I think this was a mistake, so I removed it. val_split is intentional so it generates a different dataset based on val split