 SGD in the PyTorch documentation

Pinned PyTorch2.15 primary API documentation. Algorithm pseudocode and lr parameter define the learning-rate-scaled gradient update. The scalar quadratic example sets momentum, weight decay and other modifications to zero; no PyTorch package or native build is run.

Further PyTorch2.15 locators: algorithm pseudocode and the note comparing momentum conventions, including first-buffer initialization and the handling of dampening at the first step.