Support YTA arrival observation modes in get_prob_dist_by_service

[zmek]

Feature: Support YTA arrival observation modes in get_prob_dist_by_service

Summary

Extend get_prob_dist_by_service so yet-to-arrive bed-demand PMFs can be built with observation_mode="arrived_in_window" (and optionally arrived_and_admitted_in_window) using an inpatient_arrivals frame — the same helper already used for ED-current and departures.

Simplify notebook 4d by replacing its hand-rolled YTA PMF loop with this API.

Problem

The evaluate layer supports arrived_in_window with inpatient_arrivals, but get_prob_dist_by_service does not:

• validate_observation_mode_for_component only allows admitted_at_some_point and admitted_in_window for component="arrivals".
• Those modes require ed_visits, which is wrong for YTA and direct-admission flows.

Notebook 4d works around this with nested loops calling yta_model.predict and count_observed per specialty and date, while ED-current PMFs in the same section already use get_prob_dist_by_service.

Proposed changes

aggregate.py — add inpatient_arrivals parameter and extend frame requirements:

component	observation_mode	Required frame
arrivals	admitted_at_some_point, admitted_in_window	ed_visits
arrivals	arrived_in_window, arrived_and_admitted_in_window	inpatient_arrivals
departures	departed_in_window	inpatient_visits

Update validate_observation_mode_for_component in observations.py accordingly. Pass inpatient_arrivals through to count_observed for observed counts in returned leaves.

For YTA-only FlowSelection (include_ed_yta=True, include_ed_current=False):

• Do not require ed_visits.
• Do not require curve parameters x1–y2 unless parametric in-window weighting is active.
• Keep using build_service_data / DemandPredictor so predicted PMFs match production.

Notebook 4d — replace the hand-rolled YTA block in section 2b with:

flow_sel_ed_yta = FlowSelection.custom(
    include_ed_current=False,
    include_ed_yta=True,
    include_non_ed_yta=False,
    include_elective_yta=False,
    include_transfers_in=False,
    include_departures=False,
)

ed_yta_by_specialty = get_prob_dist_by_service(
    ed_visits=None,
    snapshot_dates=eval_snapshot_dates,
    prediction_time=prediction_time,
    models=service_models,
    specialties=specialties,
    prediction_window=prediction_window,
    flow_selection=flow_sel_ed_yta,
    component="arrivals",
    observation_mode="arrived_in_window",
    inpatient_arrivals=eval_inpatient_arrivals_df,
    x1=x1, y1=y1, x2=x2, y2=y2,
    verbose=False,
)

Acceptance criteria

• arrived_in_window works on get_prob_dist_by_service with inpatient_arrivals and without ed_visits.
• admitted_in_window still requires ed_visits (regression).
• Returned leaves have correct agg_predicted and agg_observed for a minimal YTA fixture.
• Notebook 4d section 2b uses the new API; hand-rolled YTA loop removed.
• Tests added in tests/test_aggregate.py; CI green.