Add Stage 1 dataset build specs

[anth-volk]

Fixes #1036

Summary

• Add canonical Stage 1 dataset-build substep and artifact specs under policyengine_us_data.build_datasets.
• Generate Modal SCRIPT_OUTPUTS and Stage 1 contract outputs/substages from the shared specs.
• Add focused tests for artifact inventory, skip flags, step-manifest alignment, and pipeline-doc metadata.

Validation

• ruff check modal_app/data_build.py modal_app/step_manifests/specs.py policyengine_us_data/build_datasets/__init__.py policyengine_us_data/build_datasets/artifacts.py policyengine_us_data/build_datasets/specs.py policyengine_us_data/stage_contracts/dataset_build.py tests/unit/test_build_dataset_specs.py tests/unit/test_dataset_build_stage_contract.py tests/unit/test_modal_data_build.py tests/unit/test_pipeline_docs_extractor.py
• ruff format --check modal_app/data_build.py modal_app/step_manifests/specs.py policyengine_us_data/build_datasets/__init__.py policyengine_us_data/build_datasets/artifacts.py policyengine_us_data/build_datasets/specs.py policyengine_us_data/stage_contracts/dataset_build.py tests/unit/test_build_dataset_specs.py tests/unit/test_dataset_build_stage_contract.py tests/unit/test_modal_data_build.py tests/unit/test_pipeline_docs_extractor.py
• uv run --no-sync pytest tests/unit/test_build_dataset_specs.py tests/unit/test_dataset_build_stage_contract.py tests/unit/test_modal_data_build.py tests/unit/test_pipeline_doc_guards.py tests/unit/test_pipeline_docs_extractor.py tests/unit/test_pipeline_status.py tests/unit/test_stage_contracts.py tests/unit/test_step_manifest.py
• uv run --no-sync --with pyyaml python scripts/extract_pipeline_docs.py --json /private/tmp/stage1-pr1-docs/pipeline_map.json --api-json /private/tmp/stage1-pr1-docs/pipeline_api.json --markdown /private/tmp/stage1-pr1-docs/pipeline-map.md
• uv run --no-sync --with pyyaml python scripts/run_quality_guards.py
• make lint

Review Notes

• Consolidated review found no high/critical in-scope findings.
• Plain uv run ... attempted a full sync and failed on macOS x86_64 because torch==2.9.1 has no compatible wheel; validation was rerun with --no-sync per repo docs.

[MaxGhenis]

my understanding of the full pipeline:

1. download
2. stack cps, clear values that the puf could have
3. impute the missings from the puf (missings cleared in step 2 + missings from cps itself) [synthesize high income / forbes somewhere around here?]
4. impute from other sources like scf, sipp, acs
5. assign census block (and other geos from there)
6. calculate tax/ben from pe

7a. calibrate legacy

7b. calibrate new

8. publish