Function request: sample every N images as in A111

Right now the only way to get sample generation directly in the training process is to create a text file with the prompt. This is not convenient when fine tuning a model using millions of images for training. When training an embedding or hypernetwork in A111, sampling is done using "random" image data for training. For example, if it is set to create an image every 100 training steps, then images will be created from the cues for the hundredth, two hundredth, three hundredth image, and so on. This allows you to see more accurately the result of the training itself rather than a successful or unsuccessful choice of prompt.

Start by locating the training-time sampling path and the existing text-file prompt handling, then compare the described A111 behavior for embedding and hypernetwork training. Done means an optional interval generates samples from training image cues at each Nth step without requiring a prompt file, covering the requested fine-tuning workflow.

Written by the indexing model from the issue text.