Allow more granular control over which timesteps are chosen during training

Timesteps are randomly chosen during training within a specified range, defaulting to 0 to 1000. You can only adjust min and max range for timesteps chosen. However, because the min and max range apply to all training steps/epochs, it would only result in undertrained loras if anything besides the default settings are used. It would be better if there's a piecewise function parameter that allows you to choose a timestep for each step similar to the piecewise function scheduler for the learning rate. This would help with generalization and reduce overcooking since not all timesteps need to have the same emphasis.

The issue names no files or tests; start by locating the training timestep selection and the existing piecewise learning-rate scheduler. Compare their interfaces and define how a per-step timestep schedule should work, with completion demonstrated by configurable timestep choices across training steps without relying on one global range.

Written by the indexing model from the issue text.