Support end of training in the middle of an epoch

For very large graphs, training one epoch may take a very long time. People may want to define the number of total iterations for a job instead of the number of epochs.

No file, test, or entry point is named. Start by locating the training configuration and epoch-based training loop, then trace how progress and stopping are handled. Done means a job can specify total iterations, stop in the middle of an epoch, and preserve the existing epoch-based behavior.

Written by the indexing model from the issue text.