TT-NN scaled-dot-product attention decode API

Generated API page for ttnn.transformer.scaled_dot_product_attention_decode on the 'latest' docs track, read 2026-10-17. A living page with no pinned revision, so the constraints are the state of latest on that date. The page states, in its own words, that the implementation is Flash-Decode and currently only supports MQA while decoding a single token; that the op parallelizes over the batch and the K/V sequence dimension; that a batch position of -1 skips compute for that index; and that the K sequence length must be a multiple of k_chunk_size, which unpaged decode defaults to min(512, the largest power-of-two divisor of the sequence length) while paged decode uses 32 when no program config is given. The same MQA sentence appears on the paged sibling page.