v0.9.0

What's Changed

• docs(readme): mention cosmos3 policy and link design docs by @WilliamYue37 in #442
• feat(robocasa): default eval scene split to pretrain by @shuheng-liu in #443
• feat(pi05_mem): interleaved MRoPE as default position embedding by @akshay18iitg in #444
• feat(cosmos3): single-layer reasoner conditioning + backbone truncation by @WilliamYue37 in #445
• feat(pi07_paligemma): AR discrete-action decode + goal-frame eval infra by @shuheng-liu in #447
• docs(cosmos3): note on DeepSpeed ZeRO-2 vs ZeRO-3 for training by @WilliamYue37 in #446
• fix(robocasa): promote multi-env Sync eval to Async for correct renders by @shuheng-liu in #450
• docs: document benchmark_inference.py + cosmos3 latency by @WilliamYue37 in #451
• chore: bump version to 0.9.0 by @WilliamYue37 in #452

Full Changelog: v0.8.0...v0.9.0