Set up dev environment + add AGENTS.md cloud notes

[Cruiz102]

Summary

Sets up the development environment for MASK_ML and documents non-obvious setup/run caveats for future agents. No application code is modified.

The only file change is a new AGENTS.md. The environment itself is provisioned by the startup update script (registered separately), which creates a .venv, installs CPU PyTorch + requirements.txt + pycocotools + pip install -e ..

What was verified

• Dependencies install cleanly into .venv (CPU torch 2.13.0, hydra, opencv, scikit-learn, pycocotools, mask_ml).
• tests/vit_test.py passes 3/3.
• Hello-world: the real Hydra + MNIST + ImageAutoEncoder + MSE + AdamW training pipeline runs and loss decreases (4.21 → 1.38 over 100 steps).

Notes / caveats captured in AGENTS.md

• CPU-only box: GPU-only demo/ scripts and the Docker flow can't run here; PyTorch is the CPU build.
• pycocotools is required by mask_ml/utils/datasets.py but is commented out in requirements.txt/setup.py, so the update script installs it explicitly.
• torchvision's default MNIST mirror (yann.lecun.com) is blocked here and errors before the working S3 mirror; pre-place the gz files in datasets/MNIST/raw/.
• Network egress is intermittently flaky at pod start (TLS resets); network ops are retried.
• Pre-existing bug on main: eval.py imports mask_ml.model.segmentation_auto_encoder, which does not exist, breaking train.py, eval.py, and tests/auto_encoder_test.py at import time. Left unmodified per scope; flagged for the maintainer.