Default publication calibration to A100

[MaxGhenis]

Summary

• Default full publication calibration fits to A100-40GB instead of T4 for both regional and national weights.
• Update the workflow-dispatch fallback and Makefile defaults to match.
• Add a source-level contract test so the production default does not drift back to T4.

Why

The current 2024 publication package has 20,680 targets x 5,159,570 records. Both T4 fits failed before epoch 1 with CUDA OOM:

• regional T4 call fc-01KS2TFYAN0VG2985M3ED4WRNH
• national T4 call fc-01KS2TFYDDSZKRQFWJ36EPCEKQ
• failure: tried to allocate another 3.65 GiB with only 3.33 GiB free on a 14.56 GiB T4

The A100-40GB rerun from the same saved package passed that point and reached epoch 100+ for both fits.

Local checks

uv run pytest tests/unit/test_pipeline_source_contracts.py -q
uv run ruff check modal_app/pipeline.py tests/unit/test_pipeline_source_contracts.py
uv run ruff format --check modal_app/pipeline.py tests/unit/test_pipeline_source_contracts.py
git diff --check upstream/main..HEAD