Move all randomness to data package for deterministic country package

[MaxGhenis]

Summary

Move ALL random number generation from policyengine-us into dataset generation, making the country package a purely deterministic rules engine. Adds name-based seeding for order-independent reproducibility and covers all 10 stochastic programs.

MERGE ORDER: This PR must be merged BEFORE PolicyEngine/policyengine-us#7326

Background

This PR and its companion (policyengine-us#7326) supersede the earlier policyengine-us#6635, which was built against an older data package design (shared-seed RNG, no WIC draws, no SSI, uniform Medicaid rate). Rather than rebasing that branch through months of conflicts, we started fresh on current main to incorporate name-based seeding, state-specific Medicaid, and Section 1931 deprivation rules in one shot.

Changes

Name-based seeding

• Ported _stable_string_hash from policyengine-core to utils/randomness.py
• Each variable gets its own RNG via seeded_rng(variable_name), making results independent of generation order
• seeded_rng supports an optional salt parameter for per-subgroup streams (e.g., congressional districts), while the unsalted default remains distinct from any salted variant

State-specific Medicaid takeup rates

• Replaced uniform 93% rate with per-state rates (53% Utah to 99% Colorado/DC/Indiana)
• 51 jurisdictions (50 states + DC) from MACPAC enrollment targets

New variables generated

• meets_ssi_resource_test (bool, rate=0.4)
• wic_takeup_draw (float [0,1))
• wic_nutritional_risk_draw (float [0,1))

Existing variables updated

• EITC, DC PTC, SNAP, ACA, Medicaid, Head Start, Early Head Start now use per-variable seeded RNG

Related PRs

• policyengine-us companion: Move all stochastic randomness to data package for deterministic country package policyengine-us#7326
• Supersedes: Make country package purely deterministic - read stochastic variables from dataset policyengine-us#6635 (closed), Add state-specific Medicaid takeup rates and Section 1931 deprivation rules policyengine-us#7317 (fixes #7316)
• UK pattern: Move all randomness to data package for deterministic country package policyengine-uk-data#203, Make country package purely deterministic - read stochastic variables from dataset policyengine-uk#1355

Test plan

• pytest passes for stochastic variable tests
• Verified takeup rates match YAML parameters (~0.672 for ACA, etc.)
• Name-based seeding produces identical results across runs
• Salt parameter produces distinct but reproducible streams

[baogorek]

I haven't forgotten about this. Will get to it (to review) Tues, Dec 16th.

[baogorek]

@MaxGhenis apologies for taking so long on this review.

My first thought is, do we really want to add the responsibility of dealing with parameters (the take up rates) in this package when they're already handled in policyengine-us? That's a lot of yaml files if we don't need them, and it feels like duplication.

I'm not crazy about the order of the computations mattering. There will almost certainly be breaks in reproducibility if the generator is used in between these take-up decisions.

[MaxGhenis]

parallel to the -us pr, lets change these to bools

[baogorek]

FYI

❯ Do you think we did the right thing? Do you regret anything you coded? Should we have pushed back? 

● Honest take: the change is directionally right but has a couple of warts.

  Why Max is right: The float draws were a half-measure. The whole architecture says "data package owns randomness, country package is deterministic." But
  with draws, the country package was still doing draw < rate[category] - it was half-owning the randomness. Now would_claim_wic is just a data-read bool,
  same as every other takeup variable. That's clean.

  What I don't love:

  1. Parameter duplication. The WIC rates now live in both repos. If USDA publishes 2023 rates, someone has to update them in two places. The other takeup
  variables (SNAP, EITC, etc.) don't have this problem because their rates were always only in the data package. We could have deleted the parameter files
  from policyengine-us since nothing references them anymore - but that's a separate PR concern.
  2. The category gate in is_wic_at_nutritional_risk. With draws, NONE people got 0 < 0 = False for free - the rate structure handled it naturally. Now with
  default_value = True, we need an explicit category != NONE check. It works, but it's compensating for a thing the draw pattern handled more elegantly.

  Should we have pushed back? No. Consistency across all takeup variables is worth more than the cleverness of the draw pattern. But I'd flag the parameter
  duplication in the PR description - long term, the rates in policyengine-us are dead code now and should probably be removed.

I am baogorek and I've taken over the code writing