feat: Add Modal training runner

[homanp]

What does this PR do?

Adds a Modal runner for remote Gemma probe training and guarded generation on an A100 GPU. The runner persists Hugging Face downloads in open-constitution-hf-cache, saves probe outputs in open-constitution-outputs, uses a Modal huggingface-secret for HF_TOKEN, and documents the workflow with a Mermaid diagram.

Also updates the Transformers loader to use dtype= instead of deprecated torch_dtype=, adds modal as a dev dependency, and ignores local .env files.

Fixes #

Checklist

• I tested my changes
• I reviewed my own code

Validation performed:

• uv run ruff check .
• uv run ruff format --check .
• uv run pytest
• uv run python -m compileall activation_probe_mvp scripts tests modal_train.py
• Modal guarded-generation smoke test using saved probe volume: completed without pause

[open-cla]

Contributor License Agreement

All contributors are covered by a CLA.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.