Fix bf16 AMP detection for T4 in train_production.py

Fix conv2d bf16 crash on T4: colab_train_iris.py

Fix conv2d bf16 crash on T4: iris/model.py

Fix conv2d bf16 crash on T4: iris/blocks.py

Fix conv2d bf16 crash on T4: iris/pde_ssm.py

Update README with Colab training instructions and VRAM budget

Add one-click Colab training notebook with real Pokemon dataset

Fix: register context_proj as proper nn.Module (was lazy, not saved in checkpoints)

Upload iris/test_all.py

Upload iris/train_production.py

Upload iris/train.py

Upload iris/flow_matching.py

Upload iris/configs.py

Upload iris/model.py

Upload iris/core.py

Upload iris/blocks.py

Upload iris/__init__.py

Upload iris/pde_ssm.py

Upload README.md

initial commit