Mini-batch training with learning-rate schedule

Add mini-batch updates and learning-rate decay to scratch training.
Compare convergence behavior and final metrics against full-batch scratch baseline.

Start by locating the scratch-training notebook and the full-batch baseline it uses. Implement the requested mini-batch updates and learning-rate decay, then compare convergence behavior and final metrics against the baseline; the work is done when those comparisons are available.

Written by the indexing model from the issue text.