Remove the need to define a split when running predict

[drewoldag]

The current code expects that a split will be defined in the [predict] table in the config file. The user can then set the specified split to be 1.0 (100%) of the data in the data directory.

But this feels like an odd requirement for the user to update the [data_set] split definitions in order to run inference.

I would advocate for an approach that doesn't require the user to specify a split in the [predict] table, and just runs all the data found through the trained model.

[aritraghsh09]

This will be the expected default user behavior in unsupervised learning scenarios.

In supervised scenarios, users might want to run predict only on the "test" set (i.e., data - train - validation)

I don't think there is a "right" answer here.

[drewoldag]

Proposing a solution - what is the default value for the split value in [predict] is false, which would indicate "no split, just use all the data in the data directory". And if the value is not "false" then we would expect it to be one of ["train", "validate", "test"].

I think that would suppose both use cases relatively easily:

1. I just want to use an existing trained model with a bunch of data - no mods required to the config file.
2. I want to test my trained model on a subset of my input data - the user would expect to have to define the subset in the config.