🛠️ Boogu-Image-0.1 — AoTI compile

Captures a repeated transformer block's inputs from a real 1-step run, exports + AoTI-compiles it, and uploads <BlockClass>/package.pt2 to the matching repo. Pick model, stream (single = bulk 32-34 blocks, double = 2 joint-attn blocks), and whether to use max_autotune.