Create a more standard training loop interface for pretraining

[jason-fries]

Currently, clmbr_train_model conceals more familiar training loops structure from users. In most demos and APIs, the boilerplate looks like what's outlined here https://github.com/PyTorchLightning/pytorch-lightning with this structure

dataset = MNIST(os.getcwd(), download=True, transform=transforms.ToTensor())
train, val = random_split(dataset, [55000, 5000])

autoencoder = LitAutoEncoder()
trainer = pl.Trainer()
trainer.fit(autoencoder, DataLoader(train), DataLoader(val))

basically the form

• dataloader
• data splits
• model architecture
• training

Specific details around the loss are configured in the model architecture and the trainer class handles stuff like progress bars, choice of optimizer, etc.

What is the lift required to provide a demo and refactor to support this type of workflow?

[woffett]

The refactor PR puts pre-training into this kind of API:

model = CLMBRFeaturizer(config, info)
dataset = PatientTimelineDataset(extract_path, ontology_path, info_path)
model.fit(dataset)

The original clmbr_train_model still works, just uses this API. I'll leave this issue open until piton_private is updated to reflect these changes.