Simulation(dataset=DataFrame) fails when a household or benefit unit has more than one person #2266

What happens

Simulation(dataset=<pandas DataFrame>) raises ValueError when any benefit unit or household has more than one person and the frame holds a benefit unit or household column (including the identifier columns it needs).

Reproduction

On main (2.125.2, policyengine-core 3.32.27):

import pandas as pd
from policyengine_uk import Simulation

frame = pd.DataFrame(
    {
        "person_id__2025": [1, 2],
        "person_benunit_id__2025": [1, 1],
        "person_household_id__2025": [1, 1],
        "benunit_id__2025": [1, 1],
        "household_id__2025": [1, 1],
        "age__2025": [40, 38],
        "council_tax__2025": [1_500.0, 1_500.0],
    }
)
Simulation(dataset=frame)

ValueError: Unable to set value "[1 1]" for variable "benunit_id", as its length is 2 while there are 1 benunits in the simulation.

The same happens for a frame exported with to_input_dataframe() from a simulation built on a dataset with multi-person households (Unable to set value ... for variable "rent", as its length is 12 while there are 6 households).

Cause, read in the code

Simulation.build_from_dataframe (policyengine_uk/simulation.py) takes a frame with one row per person and columns named variable__period, the format to_input_dataframe() exports. It builds the entities with build_from_ids, then calls self.set_input(variable, time_period, df[column]) for every column. For a benefit unit or household variable the column still has one value per person, so its length is wrong unless every benefit unit and household has exactly one person.

policyengine-core's own loader for this format takes each group's value from its first member (population.value_from_first_person).

A second, smaller problem in the same function: get_first_array finds a column with df.columns.str.contains(variable_name + "__"), so household_id__ also matches person_household_id__2025 and benunit_id__ matches person_benunit_id__2025. build_from_ids passes the result through np.unique, so today either column gives the same entities, but the match should be exact.

Suggested fix

For a variable whose entity is not the person, set population.value_from_first_person(values), and match column names exactly (column.split("__")[0] == variable_name). Add a test that a frame with a multi-person household builds and that a dataset-built simulation's to_input_dataframe() loads back with the same results.

Found while fixing #2254 (PR #2265 does not depend on it).