Implement blockwise attention with an online softmax

Write attention that streams key/value blocks with a running maximum and normalizer, verify it against a naive reference, then study a fused kernel.

The exercise happens on Self-directed (PyTorch, optionally Triton). Canonical link: https://triton-lang.org/main/getting-started/tutorials/06-fused-attention.html. Link check: link responded (5 Oct 2026). Opened from this app 0 times; opening is not counted as an attempt.

Supports: The PyTorch version runs on CPU; the kernel part needs a GPU. Prepares derivation-level FlashAttention questions.