Add Modal long-run projection builder

[MaxGhenis]

Summary

• add a large-memory Modal entrypoint that runs the existing long-run production projection script
• stream subprocess logs and commit output artifacts to a Modal volume even when the subprocess fails
• keep HF staging upload opt-in and add fake-Modal unit coverage for command forwarding, run ID sanitization, source SHA resolution, and staging prefix reporting

Context

GitHub hosted runners are failing even one-year long-run sentinel builds with SIGTERM/no artifacts. This keeps the same production script and metadata path, but runs it on a durable large-memory Modal worker.

Tests

• uv run --locked pytest tests/unit/test_long_term_modal_projection.py -q
• uv run --locked pytest tests/unit/test_long_term_modal_projection.py tests/unit/test_long_term_calibration_contract.py -q
• uv run --locked ruff check modal_app/long_term_projection.py tests/unit/test_long_term_modal_projection.py
• uv run --locked ruff format --check modal_app/long_term_projection.py tests/unit/test_long_term_modal_projection.py
• uv run --locked python -m py_compile modal_app/long_term_projection.py
• uv run --with modal python -c "import modal_app.long_term_projection as m; print(m.app.name)"
• uv run --with modal modal run modal_app/long_term_projection.py --help
• git diff --check