Segmented-national follow-ups from the #637 review

Remaining items from the PR #637 code review, deliberately deferred:

Efficiency (review finding 9):

• Lazy-import the reduce chain inside the runner so cold run_simulation containers don't pay ~19s importing pandas/microdf/policyengine.core before routing (regional/UK requests pay it today and never use it)
• Add the POLICYENGINE_SKIP_COUNTRY_IMPORTS setdefault guard to segmented_national_reduce.py — a bare pytest tests/test_segmented_national*.py process currently imports both country models (~80s) to run pure-mechanics unit tests

Robustness (review finding 10):

• National requests with a custom data/data_version currently fan out; small custom datasets make every child fail the >100-household assertion where a monolithic run succeeded. Either add custom-data requests to the monolithic fallback, or make the partition dataset-aware.

DRY leftovers (finding 14, partially addressed):

• The poll/backoff loops in budget_window_scheduler.py and segmented_national.py remain separate (payload building + backoff step now share src/modal/fanout.py); extracting a shared poll-with-backoff helper is still open, as is consolidating the three in-repo Modal fakes (test_budget_window_scheduler.py mocks vs test_segmented_national.py fakes) into a conftest fixture

Validation debt:

• include_cliffs segmentation: verify cliff-variable reconstruction through the reduce, then lift the monolithic fallback
• Labor-supply-response segmentation: requires recomputing national household_income_decile at reduce time (the transported column is region-group-relative) and validating against monolithic — see the LSR_PARAMETER_PREFIX comment in segmented_national.py
• Budget-window children remain pinned segmented=false; enabling nested fan-out (years x 21 containers) needs capacity planning against the 100/300 pools

Related: #638 (deprecating the region:"us" national spelling).

Start with segmented_national.py and segmented_national_reduce.py, then review budget_window_scheduler.py, src/modal/fanout.py, and the related tests named in the issue. Run the segmented-national tests first and validate each deferred item against monolithic results, custom-data behavior, import time, and pool capacity; the work is done when the listed review and validation gaps are resolved.