build_from_populations keeps _user_input_keys for the holders it replaces #605

Simulation.build_from_populations replaces every population, and so every holder, but keeps Simulation._user_input_keys. After a rebuild, the record of set_input values still lists (variable, branch, period) entries for values the new holders never received. This is the same drift as #559, through a path #561 did not cover. #561 keeps the record in step for Holder.delete_arrays and clone, and subsample resets it before rebuilding.

Reproduction (core only, 3.32.27; same code path in 3.33.0)

from policyengine_core import periods
from policyengine_core.country_template import CountryTaxBenefitSystem
from policyengine_core.simulations import SimulationBuilder

FEB = periods.period("2018-02")
tbs = CountryTaxBenefitSystem()
households = {"h": {"parents": ["a"]}}
sim = SimulationBuilder().build_from_entities(
    tbs,
    {
        "persons": {"a": {"salary": {"2018-02": 1000}, "disposable_income": {"2018-02": 5000}}},
        "households": households,
    },
)
# Rebuild the same simulation with new populations and no disposable_income input.
sim.build_from_populations(tbs.instantiate_entities())
SimulationBuilder().build_from_entities(
    tbs,
    {"persons": {"a": {"salary": {"2018-01": 1000, "2018-02": 1000}}}, "households": households},
    sim,
)
print(sim.get_holder("disposable_income")._get_array_from_storage(FEB))  # None
print(("disposable_income", "default", FEB) in sim._user_input_keys)       # True
sim.calculate("disposable_income", "2018-02")  # the formula value is cached under that record

Output on 3.32.27:

after rebuild: stored 2018-02: None
after rebuild: record has disposable_income 2018-02: True
calculated disposable_income 2018-02: [1430.0]
to_input_dataframe exports disposable_income: False ['salary__2018-01', 'salary__2018-02']

to_input_dataframe does not export the stale entry. Code that reads the record directly does see it, though. PolicyEngine-UK's supplied-input helpers (policyengine_uk/utils/supplied_inputs.py) treat a recorded (variable, branch, period) with a stored value as supplied. After a rebuild, a value carried over or calculated into that period passes as a supplied input. PE-UK reproduced this for a Class 4 trading loss: profits came out at £20,000 and NIC at £445.80, where they should be £30,000 and £1,045.80. PE-UK keeps a sweep in Simulation.calculate for this case (PolicyEngine/policyengine-uk#2260); it can go once core fixes the rebuild.

PE-UK's Simulation.build_from_situation calls build_from_populations(tax_benefit_system.instantiate_entities()) and then SimulationBuilder.build_from_dict(..., self). Its property tests rebuild cloned template simulations this way (test_state_pension_age.py, test_state_pension_credit_qualifying_age.py).

Suggested fix

In build_from_populations, reset _user_input_keys to an empty set, as subsample does. Every holder is new at that point, so none of the old entries can name a stored value. Also reset _user_input_contexts if no set_input can be running across a rebuild. Add a regression test: set an input, rebuild without it, calculate that period, and check that the record no longer names it.